Professional Documents
Culture Documents
Editors:
Edited by
Kevin D. Hoover
University of California
Davis, California
"
~.
Contributing Authors ix
Acknowledgments xi
1
The Problem of Macroeconometrics
Kevin D. Hoover
2
Recent Advances in Solving and Estimating Dynamic
Macroeconomic Models 15
Beth Fisher Ingram
Commentary 47
Adrian Pagan
3
The Economics of VAR Models 57
Fabio Canova
Commentary 99
Masao Ogaki
4
Progressive Modeling of Macroeconomic Time Series:
The LSE Methodology 107
Grayham E. Mizon
Commentary 171
Charles H. Whiteman and Jon Faust
v
vi CONTENTS
5
The Econometrics of the General Equilibrium Approach to
Business Cycles 181
Finn E. Kydland and Edward C. Prescott
Commentary 199
Gregor W. Smith
6
Causal Orderings 211
Stephen F. LeRoy
Commentary 229
Clive W.J. Granger
7
On Policy Regimes 235
Stephen F. LeRoy
Commentary 253
William 1. Roberds
8
The Lucas Critique in Practice: Theory Without Measurement 263
Neil Ericsson and John Irons
Commentary 313
Eric Leeper
9
Rational Expectations and the Economic Consequences of
Changes in Regime 325
James D. Hamilton
Commentary 345
Glenn D. Rudebusch
10
Historical Macroeconomics and Macroeconomic History 351
Charles W. Calomiris and Christopher Hanes
Commentary 417
Steven N. Durlauf
11
Modeling Volatility Dynamics 427
Francis X. Diebold and Jose A. Lopez
CONTENTS vii
Commentary 467
Douglas G. Steigerwald
12
Dynamic Specification and Testing for Unit Roots and Cointegration 473
Anindja Banerjee
Commentary 501
Eric Zivot
13
Nonlinear Models of Economic Fluctuations 517
Simon M. Potter
Commentary 561
Daniel E. Sichel
Index 569
Contributing Authors
Anindja Banerjee, Wadham College and the Institute of Economics and Statis-
tics, Oxford University, Oxford, England.
Charles W. Calomiris, Department of Finance, University of Illinois, Champaign,
Illinois, U.S.A.
Fabio Canova, Department of Economics, University of Pompeu Farra, Barce-
lona, Spain; Department of Economics, University of Catania, Catania, Italy.
Francis X. Diebold, Department of Economics, University of Pennsylvania,
Philadelphia, Pennsylvania, U.S.A.; National Bureau of Economic Research,
Cambridge, Massachusetts, U.S.A.
Steven N. Durlauf, Department of Economics, University of Wisconsin, Madi-
son, Wisconsin, U.S.A.
Neil Ericsson, International Finance Division, Board of Governors of the Federal
Reserve System, Washington, D.C., U.S.A.
Jon Faust, Department of Economics, Massachusetts Institute of Technology,
Cambridge, Massachusetts, U.S.A.
Clive WoJ. Granger, Department of Economics, University of California, San
Diego, California, U.S.A.
James D. Hamilton, Department of Economics, University of California, San
Diego, U.S.A.
Christopher Hanes, Department of Economics, University of Pennsylvania, Phila-
delphia, Pennsylvania, U.S.A.
Kevin D. Hoover, Department of Economics, University of California, Davis,
U.S.A.
Beth Fisher Ingram, Department of Economics, University of Iowa, Iowa City,
Iowa, U.S.A.
John Irons, International Finance Division, Board of Governors of the Federal
Reserve System, Washington, D.C., U.S.A.
ix
x CONTRIBUTING AUTHORS
xi
1 THE PROBLEM OF
MACROECONOMETRICS
Kevin D. Hoover
This volume contains twelve chapters that discuss important topics in recent
macroeconometrics. The presumed audience is the practicing macroeconomist or
the student of macroeconomics who has some knowledge of econometrics but
who is not a specialized econometrician. Each chapter is written by respected
econometricians with the aim of providing information and perspectives useful to
those who wish to reflect on fruitful ways to use econometrics in macroeconomics
in their own work or in macroeconomics generally. The chapters are all written
with clear methodological perspectives and aim to make the virtues and limita-
tions of particular econometric approaches accessible to a general audience in
applied macroeconomics. Because each chapter also represents the considered
methodological views of important practitioners, I hope that they will also be of
substantial interest to technical macroeconometricians as well as to the intended
audience of macroeconomists. In order to bring out more fully the real tensions in
macroeconometrics, each chapter is followed by a critical comment from another
econometrician with an alternative perspective. The full chapters on competing
methodologies in Part I further highlight these tensions.
2 MACROECONOMETIUCS
Macroeconometrics in Perspective
econometric meet most notably in Lawrence Klein. Klein's (1947) book The
Keynesian Revolution provides a formal interpretation of the The General Theory.
Hard on its heels, Klein became the foremost developer of large-scale macro-
econometric models-in his case, the FRB-Penn-SSRC model that is used by the
Federal Reserve among others for policy analysis.
Models such as Klein's were important for many reasons. As was clear in the
formalizations of Hicks and Modigliani, the Keynesian system was, despite its
appeal to aggregates, intrinsically a general equilibrium system. Consumption,
investment, and the demand for money depended on national income and interest
rates; but income and interest rates in tum depended on consumption, investment,
and the demand for money. As systems of interdependent equations, macroeco-
nometric models could be seen as relatives of the Walrasian general equilibrium
models, interest in which was growing in the twenty-five years after the mid-
1930s. Once formalized the Keynesian model appeared to replicate on an aggre-
gate level the structural characteristics of microeconomic equations. In them, the
econometrics of the business cycle could be joined to the structural econometrics
of demand measurement. The Cowles Commission program in econometrics aimed
at synthesizing these two strands of earlier econometrics.
As in any synthesis, new problems arose. The most important were the problems
of simultaneous equation bias (Haavelmo, 1943) and identification (Koopmans,
1949; Hood and Koopmans, 1953). The simultaneous equations problem led to the
development of new systems estimators, while the identification problem encour-
aged the elaboration of theory that would underwrite the a priori restrictions that
the Cowles Commission program regarded as its solution.
The search for a priori restrictions, for the true structure underlying macroeco-
nomic relations, motivated the program of microfoundations for macroeconomics.
Keynes had already provided informal microeconomic rationales for consumption,
investment, and portfolio behavior in The General Theory. Klein (1947) made a
first attempt at deriving the principal components of the Keynesian model from
first principles. Subsequent work included derivations by Baumol (1952) and
Tobin (1956, 1958) of money-demand functions; Modigliani and Bromberg (1954)
and Friedman (1957) of consumption functions; Lipsey (1960) of the Phillips
curve (a component of the postwar Keynesian model not actually found in The
General Theory); Jorgenson (1963) of the investment function; and so forth. Each
of these was an attempt to provide microfoundations for a part of the Keynesian
model, and each was to some degree successful theoretically and empirically.
Their empirical success, however, was imperfect. Typically, theoretical purity
was compromised through the use of ad hoc lag distributions to improve the em-
pirical fit and predictive ability of such microfoundational macroeconomic rela-
tions. What is more, the equation-by-equation approach treated each aspect of the
Keynesian model as a separate optimization problem and did not integrate them
4 MACROECONOMETRICS
modeled as expecting precisely what the model itself predicts. Rational expecta-
tions are therefore model-consistent expectations and should be accurate up to a
serially uncorrelated error (that is, up to an unsystematic error). From the point of
view of the new classical macroeconomics, the important thing about Muth's
hypothesis was that it modeled expectations as part of the scheme of general-
equilibrium microfoundations.
Two aspects of the new classical macroeconomics are especially worthy of
notice. First, Lucas (1972a) shifted the analysis of policy away from single policy
actions to policy rules. Because of rational expectations, any policy, whether or
not it is intentionally based on a rule, may be divided into a systematic component
and a random component. Expectations are based on the systematic component,
and the random component is simply noise. Noise can lower the efficiency of the
economic system, but it cannot be used for policy because it is unsystematic. The
systematic component can be accounted for in the optimizing decisions of eco-
nomic agents. Policymakers cannot ignore the interaction between their policy
rules and the behavior of rational agents. They must design their policies taking
that interaction into account.
Second, Lucas (1972a) undermined the structural interpretation of estimated
econometric models. Despite the attempt to provide equation-by-equation micro-
foundations, Lucas demonstrated that rational expectations alone was enough to
force the econometrician to pay attention to general equilibrium interactions be-
tween equations. "Keynesians" had, for example, regarded the Phillips curve as a
structural relationship. Lucas showed, however, that the coefficients on lagged
inflation depend on the policy rule in place. When the rule changes, so must the
coefficients. This is Lucas's earliest statement of the so-called Lucas critique, the
classic and more general statement of which is found in Lucas's (1976) "Econo-
metric Policy Evaluation: A Critique."
Lucas's (1976) paper is perhaps the most influential and most cited paper in
macroeconomics in the last twenty-five years (or even the last fifty years). It plays
a key role in the organization of the current volume. Part I of the current volume
consists of four chapters presenting the four dominant macroeconometric method-
ologies of the present day. Each has roots that predate the Lucas critique, yet each
can be seen as encapsulating a particular reaction to the Lucas critique.
Beth Fisher Ingram's Chapter 2 reports on the current state of econometric
thinking with regard to structural modeling. In some sense, this is the natural
extension of the Cowles Commission program to seek out models that could be
identified on the basis of a priori economic theory. The Lucas critique could be
6 MACROECONOMETRICS
from the model along limited dimensions of interest. Calibrated models are then
used for policy analysis. The Lucas critique is answered, in Kydland and Prescott's
view, because up to the limits of the idealization the models have identified the
bedrock of tastes and technology that constrain individuals' optimization.
While Part I of the current volume presents competing econometric methodolo-
gies that are conditioned by the Lucas critique, the five chapters of Part II address
the Lucas critique directly. Chapters 6 and 7 by Stephen LeRoy are a pair. The
Lucas critique is fundamentally about the extent to which macroeconometric models
capture causal structure. LeRoy argues that there is substantial confusion in
macroeconometrics about the very notion of causal structure, and in Chapter 6
he seeks to clarify the issues. Chapter 7 is an application of some of the lessons
of Chapter 6 to the problem of econometric policy analysis. In a line of argument
related to early work of his own (Cooley, LeRoy, and Raymon, 1984) as well as
to Sims's interpretation of the Lucas critique reported above, LeRoy argues that
most formulations of the Lucas critique misunderstand the notion of a change of
regime and the proper interpretation of the rational expectations hypothesis.
Writing in the tradition of the LSE econometrics (see Chapter 4), Neil Ericsson
and John Irons examine the question of whether the Lucas critique matters in
practice. In Chapter 8, they provide a wide-ranging survey of the influence of
Lucas's article on the profession as well as concrete econometric illustrations of
cases in which the Lucas critique appears not to be practically important.
Starting from a premise similar to that of Sims, that policy regimes are gov-
erned by a probability distribution, James Hamilton in Chapter 9 discusses proce-
dures for modeling time-series in the face of regime switches. This is principally a
time-series, data-oriented approach. In contrast, Charles Calomiris and Christopher
Hanes argue in Chapter 10 that to take regime switching and structural change
seriously, it is critical to have a fuller structural understanding of the economy.
Such an understanding can be derived from a more explicitly historical approach.
On the one hand, historical macroeconomics draws on a variety of sources of
information that help to make econometric approaches to time-series data more
informative. On the other hand, history provides us with examples of alternative
regimes and clear structural change that might allow us to ascertain the importance
of the Lucas critique or the relevance of particular macroeconomic theories in
practice.
While the first two parts of the current volume are motivated by the aftershocks
of the Lucas critique, the third part looks to the future. Some of the most important
developments in econometrics have yet to percolate fully to the typical applied
macroeconomist. Each of the developments reported in the three chapters in this
part involve departures from the standard assumptions of traditional econometrics.
Econometricians have usually been content to assume that the true errors in
their estimated equations are homoscedastic or capable of being made so through
THE PROBLEM OF MACROECONOMETRICS 9
Notes
I. The history of the new classical macroeconomics and of econometrics in a new classical frame-
work is discussed in detail in Hoover (1988).
2. This chapter is the only one that reprints previously published work. The strategy in compiling
this volume was to include new and up-to-date statements of the various points of view. To my
surprise, however. I could not persuade any of at least twenty prominent practitioners of the calibration
methodology to contribute a new account of calibration to the volume. Gregor Smith's commentary
on Kydland and Prescott's paper is, however, published here for the tirst time.
10 MACROECONOMETRICS
References
Baumol, Williarn 1. (1952). "The Transactions Demand for Cash: An Inventory Theoretic
Approach." Quarterly Journal of Economics 66, 545-556.
Bemanke, Ben S. (1986). "Alternative Explanations of the Money-Income Correlation." In
Karl Brunner and Allan H. Meltzer (eds.), Real Business Cycles, Real Exchange Rates
and Actual Policies (Vol. 25) (pp. 49-100). Carnegie-Rochester Conference Series on
Public Policy. Amsterdam: North Holland.
Clower, Robert. (1965). "The Keynesian Counter-revolution: A Theoretical Appraisal." In
Frank Hahn and F.P.R. Brechling (eds.), The Theory of Interest Rates. London:
Macmillan.
Cooley, Thomas F., Stephen LeRoy, and Neil Raymon. (1984). "Econometric Policy Evalu-
ation: Note." American Economic Review 74, 467-470.
Cooley, Thomas F., Stephen LeRoy. (1985). "Atheoretical Macroeconometrics: A Cri-
tique." Journal of Monetary Economics 16,283-308.
Engle, Robert F. (1982). "Autoregressive Conditional Heteroscedasticity with Estimates of
the Variance of United Kingdom Inflation." Econometrica 50, 987-1007.
Fisher, Irving. (1911/1931). The Purchasing Power of Money. New York: Macmillan.
Fisher, Irving. (1930). The Theory of Interest. New York: Macmillan.
Friedman, Milton. (1957). A Theory ofthe Consumption Function. Princeton, NJ: Princeton
University Press.
Friedman, Milton. (1968). "The Role of Monetary Policy." American Economic Review 58,
1-17.
Frisch, Ragnar. (1933). "Propagation Problems and Impulse Response Problems in Dy-
namic Economics." In Economic Essays in Honour of Gustav Cassel: October 20th,
1933. London: George Allen and Unwin.
Haavelmo, Trgve. (1943). "The Statistical Implications of a System of Simultaneous Equa-
tions." Econometrica 11, 1-12.
Hansen, Lars Peter, and Thomas 1. Sargent. (1980). "Estimating and Formulating Dynamic
Linear Rational Expectations Models." Reprinted in Robert E. Lucas, Jr. and Thomas
J. Sargent (eds.), Rational Expectations and Econometric Practice. London: George
Allen and Unwin, 1981.
Hicks, John R. (1937). "Mr. Keynes and the Classics." Reprinted in Critical Essays in
Monetary Theory. Oxford: Clarendon Press, 1967.
Hood, W.e., and Tarjalling Koopmans (eds.). (1953). Studies in Econometric Method.
Cowles Commision Monograph 14. New York: Wiley.
Hoover, Kevin D. (1988). The New Classical Macroeconomics: A Sceptical Inquiry. Oxford:
Blackwell.
Hoover, Kevin D. (1995). "Facts and Artifacts: Calibration and the Empirical Assessment
of Real-Business-Cycle Models." Oxford Economic Paper 47,24-44.
Jorgenson, Dale W. (1963). "Capital Theory and Investment Behavior." American Eco-
nomic Review 53(2), 247-259.
Keynes, John Maynard. (1939). "Professor Tinbergen's Method." Economic Journal 49,
558-568.
THE PROBLEM OF MACROECONOMETRICS 11
Tobin, James. (1956). "The Interest Elasticity of the Transactions Demand for Cash."
Review of Economics and Statistics 38, 241-247.
Tobin, James. (1958). "Liquidity Preference as Behaviour Towards Risk." Review of
Economic Studies 25, 65-86.
Tobin, James. (1969). "A General Equilbrium Approach to Monetary Theory." Journal of
Money, Credit and Banking 1, 15-29.
Tobin, James. (1980). Asset Accumulation and Economic Activity. Oxford: Blackwell.
Tooke, Thomas, and William Newmarch. (1838-1857). A History of Prices: and of the
State of the Circulation from 1792-1856 (Vols. I-VI). Introduction by T.E. Gregory.
New York: Adelphi.
I ALTERNATIVE
ECONOMETRIC
METHODOLOGIES
2 RECENT ADVANCES IN
SOLVING AND ESTIMATING
DYNAMIC, STOCHASTIC
MACROECONOMIC MODELS
Beth F. Ingram
Two of the major objectives of macroeconomic research are to explain the behavior
of aggregate economic data and to predict the effects of policy interventions.
Within the macroeconomics literature, there are two identifiable approaches to
these issues. The reduced-form method) involves specifying a statistical model for
the variables of interest, estimating the parameters of the model, and answering the
underlying question by analyzing the estimated values of the parameters or some
function of the parameters. The coherence between the model and the data is of
primary concern; theory, in general, plays a subordinate role. The structural ap-
proach, on the other hand, entails describing a theoretical model for the relevant
macroeconomic variables and analyzing the relationships implied by the model to
answer the questions of interest. An important feature of the theoretical model is
that the parameters of the model be policy invariant; the parameters are structural,
remaining fixed under hypothetical interventions. The magnitude of the roles that
measurement and observation play in the structural approach have varied greatly
over time, being central in the work of the Cowles Commission and, more re-
cently, subsidiary in the real business-cycle (RBC) literature. The point of this
essay is to discuss the second approach-the structural program in macroeconomics.
Structural analysis grew in importance through the work of the members of
the Cowles commission in the 1940s and 1950s. In criticizing the alternative
15
16 MACROECONOMETRlCS
approach, the reduced-fonn methods exemplified at that time by Bums and Mitchell
(1946), Koopmans (1947) indicated his reasons for favoring structural modeling:
the centrality in macroeconomics of questions about the effects of policy interven-
tions, and the importance of using both theory and measurement to analyze cy-
clical fluctuations. The contributions of the Cowles group, including the advocacy
of a probabilistic viewpoint about economic phenomenon and of using systems of
equations to describe behavior, provided the impetus for the development of the
simultaneous-equation model (SEM), which obtained widespread acceptance
among academics and policy makers during the 1960s and early 1970s.
A simultaneous-equation model is composed of a set of (usually linear) equa-
tions, derived from economic theory, which describe the behavior of various
sectors of the economy. Observed data are used to assign values to the parameters
of the model, given that the parameters are uniquely identified. Once the para-
meters of the model are estimated, the model can be used to explain behavior (why
does consumption rise when income'rises?) or to forecast behavior (by how much
will consumption rise if the income tax rate falls by 10 percent?). As emphasized
by Hurwicz (1962), the latter question can be answered reliably only under the
assumption that the parameters of the model are structural-invariant-with re-
spect to the proposed intervention (the change in the tax rate).
Two events precipitated a decline in academic interest in linear simultaneous-
equation models: the empirical failure of SEMs to capture the high inflation and
high unemployment of the early 1970s, and the publication of Lucas's theoretical
critique of econometric policy evaluation in 1976. Simply stated, Lucas argued
that the simultaneous equation models in use at the time were not structural in the
sense of Hurwicz. The parameters of the specified equations could be expected to
vary with changes in policy rules, thus invalidating the policy conclusions derived
from the model. More specifically, given that agents are forward looking, the
parameters of their decision rules will change with alterations in the stochastic
environment that they face. Lucas maintained that the most reasonable way to
model a fiscal or monetary policy shift was as a change in the stochastic behavior
of policy. Hence, if the researcher had estimated the parameters of decision rules
under one stochastic policy environment, then those estimates could not be used
to analyze alternative policy environments.
An outgrowth of the Lucas critique was the development of the rational expec-
tations structural modeling approach to policy analysis. As outlined by Lucas and
Sargent (1981), a model is a description of the economic environment facing the
agents in the economy, including the system for producing goods, the preferences
of the agents, and the behavior of the policy authorities. Parameters of the model
have explicit economic interpretations; in addition, they are presumed to be struc-
tural with respect to any proposed policy interventions. Agents are assumed to
maximize utility subject to a set of constraints. In most cases, today's decisions
ESTIMATING DYNAMIC MACROMODELS 17
are contingent on what the agents expect to occur in the future. Agents fonn
expectations of future variables rationally. using all infonnation available to them
and making use of the correct statistical distribution function. 2 This leads to a set
of restrictions among the equations of the model that must be accommodated in
the solution and estimation algorithms.
The mechanics of the post-Keynesian macroeconomic paradigm were first given
expression in two related papers written by Hansen and Sargent (1980). In the first
paper, the authors outlined a model of labor demand in which a competitive finn
chooses the amount of labor to hire to maximize the discounted value of expected
future profits. The model had a linear-quadratic setup so that the (linear) decision
rule for labor demand can be derived explicitly. Since labor demand depends on
a random shock to the productivity of labor, the labor-demand schedule is stochastic.
The shock is observed by the finn but not by the econometrician; statistical
methods can be used to estimate the parameters of the demand function. 3 The
companion paper (1981) described how to extend the method to a multiple equa-
tion framework.
Lucas and Sargent maintained that the key features of this approach were the
assumptions that agents are rational optimizers and that markets clear, but both of
these assumptions have been relaxed in various applications. The unifying feature
of modem macroeconomic models is the detailed specification of the behavior of
agents (including the infonnation sets of the agent and the econometrician) and of
the institutional structure of the economy.
The goal of this essay is to outline the post-Keynesian approach to structural
modeling, as envisioned in the new classical economics of Lucas and Sargent,
and to discuss the evolution of this program over the past decade. Specifically,
I discuss extensions to nonlinear models and various estimation methods. It is
probably fair to say that the methods exemplified in Hansen and Sargent have
not enjoyed, in the ensuing period, the same success that the SEM framework did
in the 1950s and 196Os. One objective of this essay is to comment on possible
explanations.
Ea , - N(O. (j~)
En - N(O, (J ;),
with E(ErrE a,) = 0 for all f, and Pa and Prconstrained to lie in the interval (0, I).
The firm is assumed to have rational expectations: it bases its decisions on all
currently available information and uses the correct probability distribution for the
variables that it needs to forecast. Here, we assume that the firm's information
set at f contains current and past values of the shocks and decision variables, I, =
{as' rs' ks_I ' S S; f}.
A solution to this model can be characterized as a decision rule for the capital
stock such that the firm maximizes the present discounted value of its profits; the
rule will be a function of the variables in I" and its functional form will depend
on the stochastic processes, {a,}::O and {r,};:o. Given the first-order nature of the
driving processes in this example, the current capital stock, k" will depend on the
current value of r, and a I and one lagged value of capital, kt_ l •
The decision rule for capital that characterizes optimal behavior for the firm can
be obtained analytically in this model. The first-order condition for the firm's
maximization problem is a second-order stochastic difference equation in the
capital stock:
[
1 - (1 + -.!.. +
13
l!.-)B
013
+ -'!"B ]E k
13
2
r r+1
= (Ii - ar
013'
- Yo)
ESTIMATING DYNAMIC MACROMODELS 19
Note that ..1. 1..1. 2 = 11f3. Now apply this factorization to the difference equation:
(I - AJB)(I - A2 B)Ek
r ,+1
= (Ii - a,-
8{3
Yo)
.
Given the values assigned to the underlying parameters of the model, Hansen and
Sargent show that Al < I < ..1. 2,
To find a unique solution to the difference equation, we need two boundary
conditions. One is provided by the specification of the initial value of the capital
stock, k- I • The other is provided by the transversality condition, which requires
that the capital stock not grow too quickly. In order to guarantee that the trans-
versality condition is satisfied, the effect of the unstable root in the difference
equation that governs the capital stock, ..1. 2, must be eliminated. Mechanically, the
forward expansion of (I - ..1. 2 B) is applied to both sides of the difference equation.
The optimal decision rule for capital derived by Hansen and Sargent is4
Yo (2.2)
The rule has this simple form because of the AR( I) assumption for the exogenous
driving processes and the simple form for the capital-adjustment-cost term. The
persistence of the capital stock is determined by AI' which depends on {3, 8, and
'11 through equation (2.1). As expected, the current capital stock responds nega-
tively to an increase in the current rental rate and positively to a favorable tech-
nology shock.
To understand the Lucas critique in this context, contrast the effect on current
and future capital of an upward movement in rr due to a large positive shock, Em
to that of a rise in rr due to an increase in its persistence (an increase in Pr)' The
two changes can be rigged so that the effect of each on the current capital stock
is the same. However, the effect on the future capital stock is quite different. The
effect of a higher than normal shock, Ern on the capital stock, kt+s' dissipates as
s becomes large. A shift in Pr permanently alters the relationship of the capital
20 MACROECONOMETIUCS
stock to rental rate shocks and results in a permanent decrease in the stock of
capital. Lucas argued that the sorts of experiments in which macroeconomists
were interested were similar to the second experiment-a change in the parameter
Pro In this case, we need to know the structure of the model: the reduced form is
not sufficient.
The solution that we have derived provides us with a mapping from the
exogenous shocks that drive the economy to the endogenous choice variables. The
researcher will likely have available a data set that includes a subset of these
variables. This data set, in conjunction with the equilibrium mapping, allows the
researcher to estimate the parameters of the model. This is the subject of the next
section.
The parameters of the model can be estimated using the decision rule in conjunction
with the known form for the data generating process for the rental rate. MUltiply
both sides of equation (2.2) by (l - PaL) and substitute the equality (l - PaL) at
=Eat into the resulting expression. Then we have the system of equations:
0][k t 2
_ ]
o 'i-2
To derive a reduced form for this system, invert the matrix on the left side of
the equation, implicitly defining K' and Ji':
O][k t
_
2
] +
o 'i-2
ESTIMATING DYNAMIC MACROMODELS 21
with the obvious definition for the matrices AI' A2, and the constants K' and p' we
can write
(2.3)
The error terms in this expression are related to Eat and Err through the expressions
Equation (2.3) forms the basis for a likelihood function (given some assumption
about the distribution of the shocks) that can be used to estimate the reduced-form
parameters AI' A 2 , K', and p', and the variance-covariance matrix of [vkt Vrr ].5
Estimating the structural parameters entails confronting two problems: the cross-
equation restrictions and identification.
A hallmark of the assumption of rational expectations is that the structure often
imposes nonlinear cross-equation restrictions on the reduced form, in addition to
the more familiar zero restrictions. Examples of the latter include the restriction
that the coefficients on the k,_, and kt - 2 terms in the rental rate equation be zero and
that only two lags of capital appear in the capital stock equation. In this example,
the single nonlinear cross-equation restriction can be derived analytically. Let V
represent the variance-covariance matrix of [Vkt vrr],let V[i,j] represent the entry
in the ith row and jth column of l/, and let AI £i, j] represent the entry in the ith
row and jth column of AI' Define A2[i, j] similarly. Then, once AI and A2 are
estimated, the elements AI(l, I] and A 2 [I, I] can be used to solve for Pa and A.I'
The quantity Pa' which must lie between zero and one by assumption, will satisfy
(2.4)
V[2, 2] = (j;.
22 MACROECONOMETRICS
To put the discussion in context, consider the following standard growth model.
A representative agent chooses consumption, c" and the capital stock, k" to maxi-
mize lifetime utility subject to a resource constraint:
limE,Wkt+s =0.
s---+oo c,+s
The transversality condition is a necessary condition for optimality and requires
that the expected value of the capital stock not grow too quickly.
An alternative form for the Euler equation is given by
with the requirement that the forecast error have conditional mean zero, E,(11,+I)
= O. Suppose that f(A,;4» is given by
24 MACROECONOMETRICS
(2.8)
The equations (2.5), (2.6), and (2.8) comprise a set of nonlinear stochastic
difference equations that jointly determine the evolution of the shocks to the
economy and the decision variables. The state of the system at time t is the vector
s, = [kt-l A,]. Given values for the parameters of the model, the solution to the
system can be characterized by a decision rule for the capital stock that depends
on the state of the system and a parameter vector: k, =g(s,; lfI). The difficulty in
finding the decision rule is that we must guarantee that both the Euler equation and
the transversality condition are satisfied.
Most algorithms for finding solutions to nonlinear stochastic models involve
approximating the decision rule by some function g(s,; lfI), then finding a lfI that
ensures that (2.6) holds. Judd (1991) and Taylor and Uhlig (1990) discuss the
various methods in more detail; here, I show how to apply them to this specific
problem. The appendix contains a partial listing of papers appearing in the eco-
nomics literature that are concerned with solution methods.
Maximizing a quadratic function subject to linear constraints is a relatively
straightforward procedure, even when the problem is stochastic. This suggests
transforming the nonlinear problem into an appropriate linear-quadratic form and
using solution methods akin to those outlined in Hansen-Sargent. The procedure
is to calculate a first-order Taylor expansion of the first-order conditions and
constraints in either the levels of the variables or the logs of the variables of the
model. Since the stochastic steady state is not known, the expansion is performed
around the deterministic steady state.
In this application, I log-linearize the system around the nonstochastic steady
state using a first-order Taylor expansion, taking derivatives of each equation with
respect to the log of the decision and state variables. This results in the following
linear difference equation (using the second form for the Euler equation):
I
c k _k a 0 -aka - (I - 8)k 0
o -
ak u- I
1-8 + aka-I [i]+ -liP
-a(a - l)k a - 1
1- 8+ ak a- 1
0 [:-'
k'_1
= [:.
0 0 0 0 -p A'_I £A'
(2.9)
where I define c, = In(c,lc), k, = In(k,lk), and At = In(A,IA). The steady state in
this model is A = I, k = «lip - I + 8)la)lI(a-ll, c = ka - 8k. The error vector
includes 71" the Euler equation forecast error, and eA,.
Implicitly defining Mo and M I , I can invert the first matrix on the left side of
(2.9) and write
ESTIMATING DYNAMIC MACROMODELS 25
initial formulation of the model: the resource constraint, equation (2.5), and the
Euler equation, equation (2.6). Substitute the expression for capital into the re-
source constraint, and substitute the resource constraint into the Euler equation to
obtain the following expression:
The goal is to find an approximate policy function, g(s,; Y/), such that this equation
is satisfied for a subset of possible states, s, E S. (Since s, E R~, it is not possible
for us to check, in finite time, whether this equation is satisfied at every point in
the state space.)
Given a functional form g(s,; ljI), such as
g(Sf; y/) = ljIo + ljI,k,_, + ljI2 A , + ljI3 k r-l A "
the objective will be to choose ljI so that (2.11) holds at a chosen set of grid
points, kj E [kLokHI, i = I, ... , N j and A j E (A Lo AH],j = I, ... ,~. Recall that s,
= [k'_IA,I = [g(s,_\; y/) A,], implying that S,+I = [k, At+\] = [g(s,; ljI) AI+ ]' and
'
g(s,+\; y/) = g(g(s,; ljI), A '+1; ljI)
= g(g(s,; ljI), A~ exp(£,+\); ljI).
If we substitute this expression into the Euler equation, we obtain an expression
solely in terms of s, and g(s,; ljI). For a given value of ljI, we calculate
g(Sj.j; y/) = ljIo + ljI\kj + ljI2Aj + ljI3kj A j (2.12)
and a residual value at each grid point using the following expression:
Since we know the distribution of the shock, £,+\, we use numerical integration to
calculate the expectation of the term in brackets.
For the correct decision rule, the residual, r j •j , is zero at all the selected grid
points. More generally, the projection of the residual error function should be zero
in all directions; this property can be used to find values for ljI. One possibility is
to fix a number of directions equal to the number of elements in ljI and to choose
ljI so that the projection of the residual function in each of those directions is zero.
ESTIMATING DYNAMIC MACROMODELS 27
In our example, 'II has four elements, so we choose four weighting functions wp'
p = 1, ... , 4, and find the vector 'II that sets the following set of four equations
equal to zero:
N, NJ
For example, if w 1(s;) =1/(N;N), the average residual will be zero for the derived
'II.
To implement this procedure, we need a method for choosing the functional
form for g(s,; 'II) and the grid points. In order to minimize numerical errors, the
most efficient procedure is to use a class of orthogonal polynomials and choose
the grid points to be the zeros of N + I,h degree polynomial in this class. For
example, one possible class of polynomials is the Chebychev polynomials, which
are defined as
Tn(x) = cos(n cos-I(x», X E [-1, IJ
or, recursively:
To(x) = 1,
T1(x) = x,
Tn..l(x) = 2xTn(x) - Tn_t(x), x E [-1, 1].
2 2
P,('II) = LL'i,jk; = 0,
j=1 ;=1
2 2
P2('II) = LL'i.jAj = 0,
j=1 ;=1
28 MACROECONOMETIUCS
2 2
P3(lfI) = L.2/i.jki A j = O.
j=1 ;=1
As the fineness of the grid increases, this method becomes increasingly accu-
rate; we are guaranteed arbitrary accuracy as we add higher-order terms to the
approximation of the decision rule and refine the grid. In this simple problem, the
method is also rapid. However. the accuracy is obtained at the cost of computa-
tional intensity and speed. For example, given two state variables and ten grid
points for each variable. the residual function must be evaluated at 100 points. If
the function is approximated using the tensor product of polynomials of degree
two, a set of nine nonlinear equations must be solved with respect to nine para-
meters.? To evaluate the conditional expectation, Gaussian quadrature can be used
to evaluate the double integral that appears in the stochastic Euler equation. In-
creasing the state space by one variable expands the number of times the residual
must be calculated tenfold, and twenty-seven nonlinear equations must be solved
with respect to twenty-seven parameters. The method becomes computationally
expensive very quickly. In addition, the method depends heavily on the accuracy
of available minimization algorithms since a system of nonlinear functions must
be solved with respect to a large number of parameters.
Albert Marcet and Christopher Sims have suggested techniques that avoid the
computational burden of the previous method. Parameterizing expectations, out-
lined in Marcet (1989), also uses functional approximation theory. Marcet's idea
is to find a function h(s,; y/) that approximates the conditional expectation that
appears in equation (2.6) and estimate lfI by nonlinear regression. More specifi-
cally, suppose we assume a function h(s,; y/) such that
With a given setting for lfI, lfIo, we can generate a data series for k, and c, from the
resource constraint and the Euler equation:
_1_[1 - 8 + aAt+1kf-IJ
C,+I
on h(s,; lfIo} using nonlinear regression methods. Given this new lfI, lfIl. generate
a new set of data; generate another lfIby running the nonlinear regression with the
ESTIMATING DYNAMIC MACROMODELS 29
new data set. We iterate until the change in VI is less than some predetermined
tolerance level.
Marcet notes that, in some applications, the sequence of VI's can become
explosive unless the starting VI, Vlo, is close to the correct value. Marcet suggests
the following updating scheme for situations in which good initial values are
unavailable:
lfI, = At1t + (1 - A)lfI,_I'
where A. is a fraction chosen beforehand by the investigator, lfI,-1 is the parameter
vector used in the previous iteration, and t1t is the parameter estimate obtained
from the nonlinear regression. A utilitarian approach is to be taken in choosing the
parameter A; the researcher should find the value that works. 8
Finally, Sims (1984) has suggested a technique called "backsolving" that de-
creases the complexity of the problem by avoiding the need to evaluate the con-
ditional expectation in the stochastic Euler equation. The method is most easily
understood in a linearized model. 9 Recall the methodology for obtaining a solution
to the model by linear-quadratic approximation. After the system is linearized, the
researcher stabilizes the linear difference equation by imposing a set of constraints
that eliminate the effect of any explosive roots in the system. The constraint also
imposes a relationship between the fundamental errors that define the distribution
of the exogenous shocks and the forecast error in the Euler equation. Data for the
decision variables and the forecast error in the Euler equation is generated by
drawing realizations of the exogenous shocks and generating the endogenous data
from the derived specification of the linear difference equations.
However, we could just as well generate a realization of the Euler equation
shock, 1]" use the mapping in reverse to generate EAt, then use the remaining
equations to generate the exogenous shock and decision variables. We will obtain
the same realization of {A, C, k I = 1, ... , T} either way. This is the intuition
behind the method, which can" be extended to the nonlinear version of the
model. First, generate a series of Euler equation errors which have conditional
mean zero. Then use equations (2.5) and (2.7) and a version lO of the constraint
C-1Ci,.][ C, k, Ii ,]' = 0 to generate the data on consumption, the capital stock and
the technology shock. Then (2.8) can be used to calculate {t:,}, if desired, although
there is no reason to believe that this series will be serially uncorrelated or even
have mean zero. Since {A,} is not generated according to (2.8), the solution that
we have derived is no longer a solution to the model as originally formulated.
"Backsolving" does not lend itself well to exercises in which the researcher
wishes to conduct a policy experiment. As the parameters of the model are changed,
the statistical properties of the backed-out shocks change. Controlling these prop-
erties is usually very important when interpreting the outcomes of perturbations
to the model. For example, if we were interested in the effect on the capital stock
30 MACROECONOMETRICS
of a higher rate of time discount, we would want to ensure that the persistence of
the technology shock was constant as we decreased /3.
Overall, the solution techniques that are meant to be used in the nonlinear
context have not been widely adopted. For most researchers, the gain that is
expected to be obtained by using one of these methods is probably outweighed by
either the presumed technical or computational cost. Abandonment of the linear-
quadratic approximation will take place only when there is a demonstration that
this approximation leads to an incorrect analysis of the model.
In order to generate data from the model or to answer particular policy questions
within the context of the model, values must be assigned to its free parameters.
Few authors in the 1980s actually used Hansen-Sargent's recommended method-
maximum likelihood-to estimate structural business-cycle models. I I Two other
methods, both easier to implement and less computationally intensive, became the
techniques of choice for assigning values to parameters in stochastic general equi-
librium models: generalized method of moments (GMM), introduced by Hansen
(1982) (see also Hansen and Singleton, 1982), and calibration, initially used in the
computable general equilibrium literature (Johansen, 1960) and applied in the real
business-cycle literature in early papers by Kydland and Prescott (1982) and Long
and Plosser (1983). Recent techniques that have been added to the toolbox are the
method of simulated moments (Lee and Ingram, 1991; Kwan and Tierens, 1992;
Duffie and Singleton, 1993) and Bayesian methods (Canova, 1993; Kwan, 1991;
and Dejong, Ingram, and Whiteman, 1994). Finally, Ingram, Kocherlakota, and
Savin (1994) suggest an alternative framework for approaching models with data.
In this section, I contrast the various methods.
C k
C- J [i, l] C- 1[i,2]
[
o 0
0 -aka - (1 - 8)k
+ 0 0 (2.13)
[
o 0
Suppose that the data set consists of a time series on aggregate consumption.
Manipulate the previous equation system to derive an expression for EM solely in
terms of consumption; find the moving average representation for A"
A, = (1 - pLrIEM ,
Finally, replace the relevant parts of the top equation in (2.13) with these expres-
sions to obtain a fairly complicated (linear) formula involving EA, and c"l2 which
forms the basis for a likelihood function under the assumption that EA , is normally
distributed. Note that the structural parameters will enter this expression in a
highly nonlinear fashion, so that the likelihood function must be maximized nu-
merically, subject to constraints on feasible values for the parameters (such as,
o < f3 < 1). Identification of the parameters may be difficult to assess.
Suppose, now, that the data set includes measures of both consumption and the
capital stock. Since the model has only one shock, the likelihood can be formed
in terms of consumption or capital. Both consumption and capital are linear func-
tions of the same shock, implying that there is some exact linear relationship
between capital and consumption; presumably, this restriction would be rejected
if tested in available data sets. The implication, of course, is that the model is
misspecified: we are trying to model a two-dimensional stochastic process (the
joint distribution of consumption and capital) by use of a model that involves a
one-dimensional shock. The resolution of this problem is to supplement the
stochastic elements of the model, either by the addition of measurement error to
the capital or consumption series or by the inclusion of additional exogenous
shocks.
The first route was taken by Altug (1989) and McGrattan (1994). In the exam-
ple model, the error vector in (2.13) would be augmented to include a second
random element that, presumably, had classical measurement error properties: it
would be i.i.d. and uncorrelated with EM' Leeper and Sims (1994), on the other
32 MACROECONONUnlUCS
hand, incorporate enough exogenous shocks in their model to avoid the implica-
tion that the observable variables form a stochastically singular system.
Maximum likelihood, as implemented by Hansen and Sargent, required that the
researcher derive the mapping from the distribution of the exogenous shocks to the
distribution of the observable variables analytically. As noted earlier, this is not
possible in most multivariate, nonlinear macroeconomic models. However, we do
have methods that allow us to generate simulated data from these models. Kwan
and Tierens (1992) suggest making use of the simulated data to estimate para-
meters by using a partial likelihood function.
Maximum likelihood estimation requires the imposition of a significant amount
of structure on the problem in the form of restrictions on the exogenous shocks.
In general, we do not observe these shocks, nor does economic theory deliver
much information about their behavior. In addition, MLE can be difficult to im-
plement in complicated models, since it involves finding the parameter vector that
minimizes a set of highly nonlinear equations, subject to side constraints on the
parameter vector. These considerations have motivated many researchers to adopt
limited-information estimation methods such as calibration and generalized method
of moments (GMM). The following section discusses some of these methods.
One common thread links the methods discussed in this section: each uses a subset
of the restrictions available from the theoretical model in estimating the model's
parameters. GMM and calibration are probably the most widely used and criti-
cized of the methods (see Kim and Pagan, 1994). GMM and calibration do not
require specification of the form of the data-generating process of the shocks,I3 nor
do they require the user to find the decision rule that is the solution to the agent's
optimization problem. All of the methods discussed in this section allow researchers
to assign values to the parameters of the model based on particular features of the
model, while neglecting other specifics of the model or the data.
set equal to the inverse of the mean of some long-tenn interest rate. The list of
papers that use calibrated parameter values is lengthy, with the advantage of the
method being its simplicity. Calibration allows researchers to gain insight into the
range of behavior that is possible in a particular model without completing a
complicated estimation exercise. The technique is not given a fonnal statistical
foundation; standard errors are not usually calculated for the estimated para-
meters. Testing the fit of the model in this framework involves a visual compari-
son of first- and second-moment properties of the data of the model to an observed
data set.
Calibration, as initially implemented, had little foundation in fonnal statistics.
Authors using this technique do not seem to think of the assigned values as
"estimates" in a statistical sense. In addition, the testing stage of the procedure is
infonnal, and an overall metric measuring the fit of the model is not presented.
This poses several problems. First, the features of the historical record that are
used to calibrate the parameter values are chosen based on the tastes of the author.
There is no objective criterion for deciding whether a particular parameter value
is reasonable or not. The calibration literature has developed in such a way that,
often, the parameters of a model will be calibrated by simply borrowing the values
used in another, perhaps very different, modeling exercise. Second, the ad hoc
nature of the testing procedure makes it difficult to decide whether a modification
of a model in a certain direction does or does not lead to improvement. Third, in
many instances, several different data sets are used to assign values to various
parameters. There is no guarantee that the data sets are consistent with each other
or with the variables of the model. 14 Finally, if we want to take the policy conclu-
sions that we derive from a model seriously, we want that model to be connected
in some rigorous fashion to the real world. It is of interest to know, for example,
the welfare cost of an expansion in the money supply for various rates of time
discount. In the end, however, we would like to know the welfare cost of inflation
under the parameterization of the model that is relevant for the country under
study.
Several authors have attempted to ground calibration on a finner statistical
footing. Gregory and Smith (1989,1991) and Christiano and Eichenbaum (1992)
treat calibration as a method-of-moments estimator. For example, a researcher
may calibrate 0 so that it is consistent with the behavior of depreciation in ob-
served U.S. data. Since real depreciation is not constant across different types of
capital, the average rate of depreciation (the sample first moment) would be used
to estimate 0. Then uncertainty about the calibrated value of 0 arises from
sampling error in estimating the mean rate of depreciation across different types
of capital. As a second example, consider the parameter a, the share of income
flowing to capital, which is often calibrated to be equal to the mean (the sample
first moment) of the historical share of income going to capital in postwar U.S.
34 MACROECONOMETIUCS
data; the mean, of course, has a sampling distribution that exhibits some degree
of dispersion. As noted by Gregory and Smith and Christiano and Eichenbaum,
method-of-moments procedures offer a procedure for calculating standard errors
for the estimates. Reporting a standard error for a is the prototypical method used
in empirical work for conveying the amount of uncertainty a researcher has about
a parameter value to the reader.
One reason to calculate standard errors for the parameters of a model is to
provide a guide for performing sensitivity analysis. When we ask a policy question
within the context of the model, we would like know whether the answer produced
by the model is sensitive to perturbations in the parameters. Standard errors give
us an idea of the range of values that are reasonable to consider. For example, we
might estimate a rate of depreciation of 0.025 in quarterly data with a standard
error of 0.005. Based on this, we might evaluate the robustness of our conclusions
to a range of values for depreciation-say, for 0 E [0.010, O.04]-under the as-
sumption that values lying more than three standard deviations from the mean are
unreasonable from the perspective of our data set.
Calibration exercises usuaIly conclude with an informal testing phase in which
statistics calculated in a real data set are compared to statistics implied by the
theoretical model. The point of this part of the exercise is diagnostic-along which
dimension is the model failing? As noted by Kydland and Prescott (1990), there
is little point in simply rejecting a model. We would like the test of the model to
produce information about how the model can be improved.
Both Gregory and Smith and Christiano and Eichenbaum suggest classical
testing procedures designed to evaluate the statistical significance of differences
between predicted and empirical moments in sets of data. These are very similar
to the tests of overidentifying restrictions developed in the GMM literature, de-
scribed below. However, in both cases, the estimation phase is separated from the
testing phase. The parameters of the model are estimated by equating one set of
model moments to data moments, and the model is tested by measuring the dis-
tance between a second set of model and data moments; uncertainty about the
parameter estimates is not accounted for in the testing stage. In other words, the
test is conditional on the given parameter setting. One can interpret procedures
that separate the estimation of the parameters from the testing of the model as
performing GMM with an inefficient weighting matrix.
Watson (1993) suggests a classical framework for model evaluation based on
second-moment properties of the data. Rather than focusing on parameter uncer-
tainty, his approach involves augmenting the variables of the model with stochastic
error in order to match empirical second moments; the larger the amount of error
required for a match, the poorer the fit of the model. The size of the error is measured
by it second-moment properties, which can be calculated only under particular
assumptions about the cross-covariance generating function of the model data and
ESTIMATING DYNAMIC MACROMODELS 35
the observed data. Assuming that this is the zero function, Watson's method pro-
vides an upper bound on the variance of the error that will be necessary to match
model and data statistics. If the upper bound is small, the model fits well. If the
upper bound is large, we reach no conclusion.
As an alternative to the classical approach, a number of authors have suggested
using Bayesian methods to analyze structural models. An important characteristic
of a standard calibration exercise is that the researcher usually displays a fair
amount of uncertainty regarding parameter values; oftentimes, the researcher will
conduct a sensitivity analysis in which the model is analyzed over a range of
parameter values. As noted in Canova (1993), DeJong, Ingram, and Whiteman
(1994), and Kwan (1991), Bayesian methods are uniquely suited for incorporating
this prior uncertainty into the analysis. The procedure involves the specification of
prior distributions over the parameters of the model under investigation, which in
tum induces prior distributions over the statistical properties of artificial data
generated from the model. The prior distribution over model parameters can be
specified based on information derived from microeconomic data, the long-run
properties of macroeconomic data, or individual intuition. The plausibility of the
model is evaluated by a comparison of the distributions of statistical moments
induced by the model to their empirical counterparts. In Canova, empirical mo-
ments are estimated with time averages of functions of the data, which has been
filtered to induce stationarity. Uncertainty about these moments arises from sam-
pling variability in the data set. DeJong, Ingram, and Whiteman, on the other hand,
take a more symmetric view of the model and data and base inference about the
properties of the actual data on a statistical model (a likelihood function); uncer-
tainty about the statistical properties of the data arises from uncertainty about the
parameters of the likelihood function.
(2.14)
36 MACROECONOMETIUCS
for a data set of length T + I consisting of consumption, c" the capital stock, k"
and output, y,. Obviously, the length of the instrument vector, p, must equal or
exceed the number of parameters to be estimated. Note that only three of the five
parameters to be estimated appear in this expression; two parameters, p and (1,
cannot be estimated using (2.14).
Define!t('I') to be the term in brackets in (2.14), where we have suppressed the
dependence of this function on the data set, and 'I' = [a f3 8]. Suppose that z" t
= I, ... , T, are vectors of length three, so that the number of instruments equals
the number of parameters to be estimated. The estimates are obtained by finding
the value of 'l'that minimizes the following function:
B =E[dr('I'o)/d'l' @ z,),
Q = E [ f ( '1'0) r('1'0)'].
The weighting matrix that yields the smallest asymptotic variance is W = .a-I,
reducing the covariance matrix to (8'Q-I B)-I. Since the weighting matrix depends
on the unknown parameter vector, '1'0' a two-step procedure can be employed:
obtain an initial estimate of 'I' by minimizing the criterion function with W = I,
employ this estimate to evaluate B and Q, then reminimize using W = Q-I.
If we expand the instrument vector so that the number of instruments exceeds
the number of parameters to estimate, the overidentifying restrictions provide a
test of the fit of the model. In that case, the minimized value of the criterion
function has an asymptotic X2 distribution with degrees of freedom equal to the
difference between the dimension of the instrument vector and the parameter
vector. Lack of fit, in this case, indicates that there is information contained in the
instrument vector that is relevant for forecasting the Euler equation shock. 15
It may not be possible to use GMM to estimate all of the parameters of a
general-equilibrium model: the parameters that appear in the stochastic Euler
equations may not comprise the entire vector of model parameters, or the para-
meters may not be identified. In addition, the method depends on having an
ESTIMATING DYNAMIC MACROMODELS 37
observed data set that corresponds to the variables appearing in the stochastic
Euler equation. This will not be the case if, for example, the model includes a
preference shock or certain types of agent heterogeneity or unobserved transac-
tions costs. Finally, as pointed out by Kocherlakota (1990) and Kim and Pagan
(1994), the quality of the asymptotic distribution of the GMM estimator depends
crucially on the correlation between the instrument vector, Z" and (af/()yJ). Lack
of an adequate instrument vector can entail a substantial departure from the as-
ymptotic distribution in the small samples typically available in macroeconomic
applications.
E
[1
= - 2,m(x
TIN
T 1=1
,) - - 2,m(y.(",»
N .=1
] [1 TIN
W - 2,m(x ,) - - 2,m(y.(",»
T 1=1 N .=1
]'
.
Using an argument similar to Hansen's, Lee and Ingram show that the estimator
has an asymptotic normal distribution with mean "'0
and covariance matrix:
where
n =NIT,
B = E[dm(Yj(",»I()yJ],
.a = var[.!.Im(x,)].
T 1=1
Choosing W= [ (I + lIn) .arl produces the smallest asymptotic covariance matrix.
As the length of the simulated series expands relative to the length of the data
38 MACROECONOMETIUCS
series, n ~ 00, the term lin becomes negligible and the asymptotic variance is
reduced.
If the number of statistics chosen exceeds the number of parameters to be
estimated, then T*E has an asymptotic X2 distribution with r - q degrees of free-
dom, providing a test of fit of the model. In this procedure, the test of fit will give
us information about how well the model replicates a set of statistics calculated
in a real data set. By altering the set of statistics, the researcher can gather infor-
mation about the dimensions along which the model is performing poorly.
This estimation procedure is straightforward, and allows the researcher to con-
centrate attention on the parts of the model that are relevant for the question at
hand. However, it depends crucially on being able to simulate data from the model
quickly. For every candidate parameter vector analyzed in the minimization algo-
rithm, the model must be solved and simulated. If a gradient method is used for
minimization, the model must be solved and simulated an additional number of
times in order to evaluate the gradient in q (the number of parameters) directions.
Lee and Ingram provide no guidance in choosing the set of moments to be used
for estimation; different sets of moments will lead, presumably, to different point
estimates. Gallant and Tauchen (1994) propose a systematic method for determin-
ing which moments to employ in a simulated-moments estimation exercise. The
idea is to make use of reliable and well-developed statistical models of the target
data set and to use moments derived from the statistical model to estimate the
parameters of the theoretical model. The statistical model need not necessarily
correspond to the likelihood function that is implied by the theoretical model.
The procedure involves using the score-the derivative of the log density-of the
statistical (or auxiliary) model. The set of moments is then the expectation of the
score under the structural model. Gallant and Tauchen show that if the auxiliary
model nests the structural model, then the estimator is as efficient as the maximum
likelihood estimator. The Gallant and Tauchen moments estimator is motivated by
statistical considerations (which empirical model provides the best statistical de-
scription of the data?) rather than economic considerations (is the forecast error
in the stochastic Euler equation zero?) as in GMM and MSM.
Bayesian Methods
The (classical) estimation methods discussed above involve the use of a data set
for the estimation and evaluation of a theoretical model. In contrast, Dejong,
Ingram, and Whiteman (1993, 1995) develop an approach, based on Bayesian
principles, in which a theoretical model is used to aid in the estimation of a re-
duced form, statistical model. The theoretical model plays the role of prior infor-
mation, which is formally combined with the statistical model-characterized by
ESTIMATING DYNAMIC MACROMODELS 39
the likelihood function-via Bayes's rule. Hence, the approach enables the fonnal
incorporation of theoretical restrictions (such as cross-equation or identifying
restrictions) into the estimation of reduced fonn models.
The procedure involves specifying a prior distribution over the parameterization
of the theoretical model, then mapping the prior into the parameter space of the
reduced-fonn model under investigation. After combining the prior with the like-
lihood function, posterior analysis over functions of interest (correlation patterns,
impulse response functions, and so on) is straightforward. In addition, it is possi-
ble to generate numerically posterior distributions over the structural parameters
of the theoretical model. Unlike classical procedures, the Bayesian posterior dis-
tributions will reflect prior information that the researcher possesses about the
values of the parameters of the model.
Ingram and Whiteman (1994) use the method in the context of a forecasting
exercise, demonstrating that a particular real business-cycle model carries infor-
mation that is useful for predicting movements in consumption, output, hours, and
investment. Dejong, Ingram, and Whiteman (1993) examine the impulse response
functions generated by a VAR that had been estimated subject to the restrictions
imposed by a monetary general equilibrium model.
With the exception of GMM and, to some extent, backsolving, all of the methods
of solution and estimation outlined above require that we specify the nature of the
exogenous shocks in the model a priori. In most applications, the number of
exogenous shocks is relatively small, and the data-generating process for vector
shock process is simple, perhaps AR(I). Model shocks, however, are unobservable;
they represent the unmeasurable parts of the model. Theory tells us little about
their behavior a priori. Hence, there is little justification for imposing structure on
these processes except as a necessary prerequisite for many methods of model
analysis.
A few authors have begun to use the structure of the model and the observed
data to make inferences about the exogenous shocks (Ingram, Kocherlakota, and
Savin, 1993, 1994; Burnside, Eichenbaum, and Rebelo, 1993; Baxter and King
1991; Parkin, 1988). The technique can be illustrated in the growth model de-
scribed above, under the assumption that a = 1 (making production linear in
capital). A solution to this optimization problem is given by
C, = (1 - /3)(1 - 0 + A,)k'_b
k, = f3(1 - 0 + A,)k'_h
y, = A,k'_l'
40 MACROECONOMETRICS
Suppose that we have data on consumption, c" and output, Y" Then a series for
the technology shocks. A" can be obtained from
A, = y,(l- {3).
c,_If3
In general, it will not be the case that both of these expressions produce the same
series for A" The problem is that the model is stochastically singular: one single
shock. A" drives fluctuations in both consumption and output. This assumption
implies that there is an exact, nonstochastic relationship between c, and Y,:
c, = f3y, + (1 - {3)(1 - 8)c'_I'
We would be quite surprised if this equation held in our data set for any values
of the structural parameters f3 and 8.
Most macroeconomists would acknowledge that variables such as consump-
tion, output, labor hours, and wages are not driven by a single stochastic shock.
A model that incorporates this assumption is misspecified and will be rejected
when subjected to a test of fit. The solution proposed by some authors has been
to simply reject formal statistical estimation and testing of models. Ingram,
Kocherlakota, and Savin (1994) argue that this resolution of the problem ignores
much of the useful information available in the data set. They propose that models
be augmented by additional shocks, until the number of shocks in the model
equals the number of observed variables in the data set. In the example, suppose
that 8 were random. Then the solution to the model is essentially the same:
c, = (1 - {3)(1 - 8, + A,)k'_I'
k, = f3(1 - 8, + A,)k'_I'
Now the equation that ties consumption to output and lagged consumption is
c, = f3y, + (1 - f3)(1 - 8,)c t _ l·
Unless we have restricted {8" t = I, ... , T} a priori, for any f3, we can always find
a sequence for depreciation such that the previous equation holds. The parameter
f3 is unidentified in this model. To identify f3, we require some additional assump-
tions about the behavior of the stochastic process, {8t}. An assumption about the
ESTIMATING DYNAMIC MACROMODELS 41
data generating process for {D,} imposes a restriction on the relationship between
current consumption, current output, and lagged consumption. This restriction can
be used to estimate {3.
If we are unwilling to impose structure on the exogenous process but are
willing to restrict the possible values of {3, the model can be used "in reverse" to
back out some information about the exogenous shocks. Given a value for {3 and
data on consumption and output, we can derive the following expressions for the
shocks:
A,({3) = (I - {3)y,l({3c,),
D,(f3) = I + (1 - {3)y,!({3cr- t) - c,!({3c,_t).
In other words, we can use the information contained in the data set to derive some
information, in the form of a realized series, for the exogenous shocks. These
backed-out shocks can then be analyzed in order to determine their relative impor-
tance in driving certain movements in consumption or output.
Ingram, Kocherlakota, and Savin have successfully applied this method in two
problems. However, in more complicated models, the mapping from observable
variables to exogenous shocks cannot be derived analytically. In addition, a re-
searcher might be interested in conducting policy experiments with the model.
This requires knowledge of the data-generating process for the shocks. This method
does not deliver this information directly, and it is not clear that there are enough
observations in available data sets to uncover this process nonparametrically.
Concluding Remarks
with more realistic institutional frameworks and agent behavior and finding
methods for solving these complicated models. Many of these solution algorithms
became feasible with the advent of high-speed desktop computers.
Many authors avoided formal statistical analysis of models, arguing that
macroeconomic models are inherently misspecified and formal statistical proof of
that fact did not add to our knowledge (Prescott, 1986; Kydland and Prescott,
1990). It was not until the end of the 1980s that the profession again turned its
attention to estimation and testing. By then, GMM had fostered a set of methods
for estimation: estimation by simulation, Bayesian methods, and a rigorous inter-
pretation of calibration. The methods outlined in this paper enhance our under-
standing of our models and their relationship to the real economy in two ways.
First, calculating a standard error for a parameter estimate can aid in sensitivity
analysis by informing us about the empirically relevant range of the parameter.
Second, analyzing the fit of the model can tell us about the dimensions along
which the model fails and the dimensions along which the model performs well.
Acknowledgments
This essay benefited from the suggestions and comments of Charles Whiteman,
Narayana Kocherlakota, N. Eugene Savin, Adrian Pagan, and Eric Leeper. The
financial support of NSF grant SES92-23257 is gratefully acknowledged.
Notes
I. Examples include the VAR, ARCH, and ECM frameworks discussed in Chapters 3, 4, 8, and
II in this volume.
2. As noted in Kim and Pagan (1994), one feature that distinguishes the rational expectations
structural modeling approach from the SEM approach is this informational assumption. In the former,
agents use all available information; in the latter, agents use a subset of the available information.
3. If the shock was observable by the econometrician, the model would imply an exact relationship
between a set of observable variables. In that case, formal estimation of the parameters is trivial.
However, unless this exact relationship were satisfied in the data for some parameter vector, the model
would be rejected.
4. This derivation is also available in Sargent (1979). Whiteman (1983) addresses the solution to
more general linear difference equations.
5. This requires an assumption about the distribution ofthe £'s. Hansen and Sargent recommended
quasi-maximum likelihood estimation. which involves using a normal likelihood in the presence of
possibly non-Gaussian errors.
6. If the number of eigenvalues that exceed (I/{3)"2 is greater than one, then there is no solution
to the model that satisfies the transversality condition; if the number is less than one, there are multiple
solutions each indexed by the initial level of consumption.
ESTIMATING DYNAMIC MACROMODELS 43
7. If x and y represent the variables, we have the constant tenn in addition to the tenns x, y, xy,
x 2, y2, xy2, x 2y, and X2y 2.
8. Judd (1991) presents a critique of this method, to which Marcet (1993) responds.
9. There is no reason to backsolve a linear model; I use this eumple for pedagogic purposes.
10. We can add any function of c" k" and A, for which!(c" k" A,,) =0 and such that the first-order
Taylor expansion is equal to C-'[i,.J[c, ~t A,)' = O.
11. The exceptions of which I am aware are Altug (1989), McGrattan (1994), and Leeper and Sims
(1994).
12. If the available data was the capital stock instead of the consumption series, we could use the
same steps to derive an equality involving k, and EM'
13. However, if a researcher is planning to perfonn a policy experiment in which the model is
simulated, or if the researcher is going to examine the statistical properties of the series in the model,
this fonn will need to be specified.
14. This point was made by Singleton (1988).
15. A more thorough discussion of this material is contained in Hamilton (1994).
References
Altug, Sumru. (1989). "Time-to-Build and Aggregate Auctuations: Some New Evidence."
International Economic Review 30, 889-920.
Baxter, Marianne, and Robert King. (1991). "Productive Externalities and Cyclical Vola-
tility." Manuscript, University of Rochester.
Bums, Arthur, and Wesley Mitchell. (1946). Measuring Business Cycles. New York:
Columbia Free Press for the NBER.
Burnside, C. M. Eichenbaum, and S. Rebelo. (1993). "Labor Hoarding and the Business
Cycle." Journal of Political Economy 101, 245-273.
Canova, Fabio. (1993). "Sensitivity Analysis and Model Evaluation in Simulated Dynamic
General Equilibrium Models." Manuscript, University of Rochester.
Christiano, Lawrence, and Martin Eichenbaum. (1992). "Current Real Business Cycle
Theories and Aggregate Labor Market Auctuations." American Economic Review 82,
430-450.
Dejong, David, Beth Ingram, and Charles Whiteman. (1993). "Beyond Calibration." Journal
of Business Economics and Statistics, forthcoming.
Dejong, David, Beth Ingram, and Charles Whileman. (1993). "Analyzing VARs with
Monetary Business Cycle Model Priors." Proceedings of the American Statistical As-
sociation, Bayesian Statistical Science Section (pp. 160-169).
Duffie, Darrell, and Kenneth Singleton. (1993). "Simulated Moments Estimation of Markov
Models of Asset Prices." Econometrica 61, 929-952.
Gallant, R., and G. Tauchen. (1994). "Which Moments to Match?" Manuscript.
Gregory. Allan, and Gregor Smith. (1989). "Calibration as Estimation." Econometric Re-
views 9, 57-89.
Gregory, Allan, and Gregor Smith. (1991). "Calibration as Testing: Inference in Simulated
Macroeconomic Models." Journal of Business and Economic Statistics 9,297-303.
Hamilton, James. (1994). Time Series Analysis. Princeton. NJ: Princeton University Press.
44 MACROECONOMETRICS
Hansen, Lars. (1982). "Large Sample Properties of Generalized Method of Moments Es-
timators." Econometrica 50, 1029-1054.
Hansen, Lars, and Thomas Sargent. (1980). "Formulating and Estimating Dynamic Linear
Rational Expectations Models." Journal of Economic Dynamics and Control 2, 7-46.
Hansen, Lars, and Thomas Sargent. (1981). "Linear Rational Expectations Models for
Dynamically Interrelated Variables." In Robert Lucas and Thomas Sargent (eds.), Ra-
tional Expectations and Econometric Practice (pp. 127-158). Minneapolis: University
of Minnesota Press.
Hurwicz, Leonid. (1962). "On the Structural Form of Interdependent Systems." In Ernest
Nagel et aL (oos), Logic and Methodology in the Social Sciences (pp. 232-239). Stanford,
CA: Stanford University Press.
Ingram, Beth, Narayana Kocherlackota, and N. Eugene Savin. (1993). "What's the Culprit?
Letting the Data Tell Us About the Great Productivity Slowdown." Manuscript, Univer-
sity of Iowa.
Ingram, Beth, Narayana Kocherlakota, and N. Eugene Savin. (1994). "Explaining Business
Cycles: A Multiple Shock Approach." Journal of Monetary Economics 34, 415-428.
Ingram, Beth, and Charles Whiteman. (1994). "Supplanting the 'Minnesota' Prior: Fore-
casting Macroeconomic Times Series Using Real Business Cycle Model Priors." Jour-
nal of Monetary Economics 34, 497-510.
Johansen, L. (1960). A Multi-sectoral Study of Economic Growth. Amsterdam: North
Holland.
Judd, Kenneth L. (199Ia). "Minimum Weighted Residual Methods for Solving Aggregate
Growth Models." Discussion Paper 49, Institute for Empirical Macroeconomics.
Judd, Kenneth L. (1991b) "Numerical methods in Economics." Manuscript.
Kim, K., and Adrian Pagan. (1994). "The Econometric Analysis of Calibrated Macroeco-
nomic Models." Manuscript, Economics Program RSSS, Australian National University.
King, Robert, Charles Plosser, and Sergio Rebelo. (1988). "Production, Growth and Busi-
ness Cycles: I. The Basic Neoclassical ModeL" Journal of Monetary Economics 21,
195-232.
Kocherlakota, Narayana. (1990). "On Tests of Representative Consumer Asset Pricing
Models." Journal of Monetary Economics 26, 285-304.
Koopmans, Tjalling. (1947). "Measurement Without Theory." Review of Economic Statis-
tics 29, 161-172.
Kwan, Yum-Keung. (1991). "Bayesian Calibration and Estimation of Business Cycle
Models." Manuscript, Graduate School of Business, University of Chicago.
Kwan, Yum-Keung, and Ingrid Tierens. (1992). "Estimating Non-Linear Rational Expec-
tations Models by Simulated Method of Partial Likelihood.' Manuscript, University of
Chicago.
Kydland, Finn, and Edward Prescott. (1982). "Time to Build and Aggregate Fluctuations."
Econometrica 50, 1345-1370.
Kydland, Finn, and Edward Prescott. (1990). 'The Econometrics of the General Equilib-
rium Approach to Business Cycles." Staff Report No. 130, Federal Reserve Bank of
Minneapolis Research Department.
Lee, Bong-Soo, and Beth Ingram. (1991). "Simulation Estimation of Time Series Models."
Journal of Econometrics 47, 197-205.
ESTIMATING DYNAMIC MACROMODELS 45
Leeper, Eric, and Christopher Sims. (1994). "Towards a Modem Macroeconomic Model
Useable for Policy Analysis." Manuscript, Yale University.
Long, J., and Charles Plosser. (1983). "Real Business Cycles." Journal ofPolitical Economy
91,39-69.
Lucas, Robert E., Jr. (1976). "Econometric Policy Evaluation: A Critique." In Karl Brunner
and Alan Meltzer (eds.), The Phillips Curve and Labor Markets. Amsterdam: North-
Holland.
Lucas, Robert E., and Thomas Sargent. (1981). "After Keynesian Macroeconomics." In
Robert Lucas and Thomas Sargent (eds.), Rational Expectations and Econometric Prac-
tice. Minneapolis: University of Minnesota Press.
Marcet, Albert. (1989). "Solving Nonlinear Models by Parameterizing Expectations."
Manuscript, Carnegie Mellon Graduate School of Industrial Administration.
Marcet, Albert. (1993). "Simulation Analysis of Dynamic Stochastic Models: Applications
to Theory and Estimation." Working Paper No.6, Universitat Pompeu Fabra Economics.
McGrattan, Ellen. (1994). "The Macroeconomic Effects of Distortionary Taxation." Jour-
nol of Monetary Economics 33, 573-601.
Parkin, Michael. (1988). "A Method for Determining Whether Parameters in Aggregative
Models Are Structural." Carnegie-Rochester Conference Series on Public Policy 29,
215-252.
Prescott, Edward. (1986). "Theory Ahead of Business Cycle Measurement." Carnegie-
Rochester Series on Public Policy 25, 11-44.
Sims, Christopher A. (1984). "Solving Stochastic Equilibrium Models 'Backwards.''' Dis-
cussion Paper No. 206, Center for Economic Research, University of Minnesota.
Singleton, Kenneth. (1988). "Econometric Issues in the Analysis of Equilibrium Business
Cycle Models." Journal of Monetary Economics 21, 361-386.
Taylor, John, and Harald Uhlig. (1990). "Solving Nonlinear Stochastic Growth Models:
A Comparison of Alternative Solution Methods." Journal of Business and Economic
Statistics 8, 1-18.
Watson, Mark. (1993). "Measures of Fit for Calibrated Models." Journal of Political
Economy 101, 1011-I04l.
AppendiX to References
The following papers deal with methods of solution for nonlinear stochastic general-equi-
librium models. The list is not exhaustive.
Baxter, Marianne, Mario Crucini, and K.G. Rouwehorst. (1990). "Solving the Stochastic
Growth Model by a Discrete State Space Euler Equation Approach." Journal of Busi-
ness and Economic Statistics 8, 19-22.
Christiano, Lawrence. (1990). "Solving the Stochastic Growth Model by Linear-Quadratic
Approximation and by Value-Function Iteration." Journal of Business and Economic
Statistics 8, 23-28.
Coleman, W. John III. (1990). "Solving the Stochastic Growth Model by Policy Function
Iteration." Journal of Business and Economic Statistics 8, 27-30.
46 MACROECONOMETIUCS
den Haan, Wouter. (1991). "An Introduction to Numerical Methods to Solve Stochastic
Dynamic Models." Carnegie-Mellon Graduate School of Industrial Administration.
den Haan, Wouter, and Albert Marcet. (1990). "Solving the Stochastic Growth Model by
Parameterizing Expectations." Journal of Business and Economic Statistics 8, 31-34.
Gagnon, Joseph. (1990). "Solving the Stochastic Growth Model by Deterministic Extended
Path." Journal of Business and Economic Statistics 8, 35-36.
Ingram, Beth. (1990). "Solving the Stochastic Growth Model by Backsolving with an
Expanded Shock Space." Journal of Business and Economic Statistics 8, 37-38.
Judd, Kenneth L. (199Ia). "Minimum Weighted Residual Methods for Solving Aggregate
Growth Models." Discussion Paper 49, Institute for Empirical Macroeconomics.
Judd, Kenneth L. (199Ib). "Numerical Methods in Economics." Manuscript.
Labadie, Pamela. (1990). "Solving the Stochastic Growth Model by Using a Recursive
Mapping Based on Least Squares Projection." Journal of Business and Economic Sta·
tistics 8, 39-40.
Marcet, Albert. (1989). "Solving Nonlinear Models by Parameterizing Expectations."
Manuscript, Carnegie Mellon Graduate School of Industrial Administration.
Marcet, Albert. (1993). "Simulation Analysis of Dynamic Stochastic Models: Applications
to Theory and Estimation." Working Paper No.6, Universitat Pompeu Fabra Economics.
McGrattan, Ellen R. (1993). "Solving the Stochastic Growth Model with a Finite Ele-
ment Method." Working Paper 514, Federal Reserve Bank of Minneapolis Research
Department.
McGrattan, Ellen. (1990). "Solving the Stochastic Growth Model by Linear-Quadratic
Approximation." Journal of Business and Economic Statistics 8, 41-44.
Sims, Christopher A. (1984). "Solving Stochastic Equilibrium Models 'Backwards.''' Dis-
cussion Paper No. 206, Center for Economic Research, University of Minnesota.
Sims, Christopher. (1990). "Solving the Stochastic Growth Model by Backsolving with a
Particular Nonlinear Fonn for the Decision Rule." Journal of Business and Economic
Statistics 8, 45-48.
Tauchen, George. (1990). "Solving the Stochastic Growth Model by Using Quadrature
Methods and Value-Function Iteration." Journal of Business and Economic Statistics 8,
49-52.
Taylor, John, and Harald Uhlig. (1990). "Solving Nonlinear Stochastic Growth Models:
A Comparison of Alternative Solution Methods." Journal of Business and Economic
Statistics 8, 1-18.
Commentary on Chapter 2
Adrian Pagan
Beth Ingram has produced a very good account of methods for' solving and
estimating small models that feature rational-expectations or intertemporal-
optimization constraints. My discussion here will not provide much in the way of
new techniques but rather is intended to raise a number of questions. Some of
these have a bearing on estimation and solution procedures, while others aim to
highlight my perception that an important "tension" in macroeconometrics is that
academics are paying insufficient attention to the needs of those who must utilize
macroeconometric models. Indeed, the latter fact has resulted in a separate litera-
ture that seeks to utilize theoretical insights but yet retain practical relevance.
Let me begin with the issue of solution. Beth's survey demonstrates very clearly
that difficulties in this arena stem from two characteristics of the class of models
she considers: nonlinear relations and the fact that the forcing variables of these
models are taken to be stochastic rather than deterministic, meaning that an expec-
tation has to be evaluated. As she observes, the most popular approach has been
to work with linearized models, a strategy for which good arguments can be made.
Foremost among these is the fact that any solution method must be approximate
and reflect a tradeoff between the accuracy available from the approximation and
the specification errors created by oversimplification of the model. There is little
point in being highly accurate if the degree of simplification needed to attain this
accuracy means that the model is of little use. As she points out in her survey, even
when there are only two state variables, many of the "contraction mapping" type
procedures become computationally infeasible, and it is hard to imagine many
realistic models possessing only two state variables. In this context it is interesting
to reflect on the interchange a few years ago at the Santa Fe Institute between
physicists and economists; the former were amazed at the desire of the latter for
exact solutions as they were more than happy with approximations. Given the
versatility of the linear approximation in handling large models, my feeling is that
we should be paying more attention to improving it, and to understanding its
limitations, rather than seeking greater accuracy with algorithms that apply only
in very restricted circumstances. 1
The computational general-equilibrium (CGE) literature faced a similar chal-
lenge in the 1980s. Following Johansen (1960) one strand of this literature, rep-
resented in Dixon et al. (1982), linearized around equilibrium solutions and then
solved for the changes in output variables caused by movements in control variables
47
48 MACROECONOMETRICS
this system are unlikely to be of the standard type-that is, some of the asymptotic
distributions of coefficient estimators will be nonnormal. If such processes are in
evidence, the application of GMM would need to be done with some care. Con-
sider, for example, the moment conditions used in her equation (2.7), writing the
population restriction as E[Z,17,+.J =O. Following the argument in Hall (1978) and,
as evident from her equation (2.9), C, will have an AR root that is close to unity-
that is, be near integrated. GMM applied to the estimation of, say, {3 using (2.7)
effectively involves the application of an instrumental variables estimator with z,
as instrument and d17,+/d{3 = (1 - 8 + ft.k,-IYt+\)!CI+ as regressor, and the latter is
likely to inherit the near-integrated properties of C,+I'' An approach that seeks to
obviate this problem is to recognize that c, is in the information set and there-
fore the moment conditions might be alternatively set up as E, [z, {I - {3(C!C,+I)
(1 - 8 + ft.k~IYt+I)} ] = 0; in this incarnation all regressors will be /(0). However,
it is not clear that this really represents a solution. To understand why, replace the
conditional expectation by the observed value, producing
Thus the GMM estimator using these moment conditions can be regarded as
coming from a regression with error term £'+1 = c,17t+I' Consequently, a question
that arises is, what are the properties of £, and 17,? If TI, was U.d.(O, 0'2) then,
although £, would be a martingale difference with respect to the information set,
it would possess very strong dependence in the second moment, due to the fact that
the conditional variance of £, depends on a near-integrated process c,. Such a pro-
cess would not fit easily into the conditions imposed by Hansen (1982) in his proof
of the asymptotic properties of GMM. Oddly enough, there has been very little
discussion of which type of moment condition is likely to be the best to use. Fun-
damentally the problem is that rational expectations only impose requirements
such as E,(£,+I) = 0, and nothing else about the nature of the random variable is
known.
Some insight is available by simulating values for £" using the method de-
scribed in Ingram's paper. 2 I fixed 0 = .25 and a = .4 and tried a number of values
for 0'= std (£A') and p. In general the var(TI,) depends on C,_I' e.g. when p = .95
and 0'= .007 the regression of 2000 values of TI;against unity and C/-I yields a t
ratio for the coefficient of C/-I of -2.6. Conditional heteroskedasticity is present
in E, as well, but it is considerably muted, with a t-ratio of -2.14. Across a range
of parameter values it seemed that E, was a better-behaved random variable than
Til' pointing to the desirability of formulating the Euler equation restrictions in
terms of changes in variables. Nevertheless, the conditional heteroskedasticity in
£, tended to remain, and might therefore might cause some concern, owing to the
fact that c, is close to an /( 1) process. The most extreme case I encountered was
so MACROECONOMETRICS
when (1= .015 and p = 1. In this instance the t ratio for the regression that has 11;
as dependent variable was -18.48, and a plot of r-lI.;=I11~, a quantity used in
Pagan and Schwert (1990) and Mandelbrot (1963) to assess the existence of
moments, showed no signs of stabilizing even when r = 2000.
There are other problems with employing GMM, as the correlation of instru-
ments with regressors in these models is frequently very low (see Pagan and
J ung' s 1993 analysis of the RBC model of Mao, 1990). When the correlation is
identically zero, Phillips (1989) pointed out that the distribution of the IV estima-
tor departed substantially from the normal distribution, and Staiger and Stock's
(1993) analysis of the near-zero correlation case showed that these effects can
persist in quite large samples. Consequently, it is not clear that estimating para-
meters from the Euler equation is as attractive as it looks.
Such reservations hint at the desirability of using maximum likelihood, but
there are other difficulties associated with this estimation option that need to be
appreciated. Ingram points to the fact that most linearized stochastic general equi-
librium (SGE) models reduce to a VAR (see equation (2.9) of her paper for
example). Such a VAR might be represented as
Z~ =Aztl + reI' (1)
assuming that the system is only a first-order one. An immediate problem is that
the number of shocks £, is generally smaller than the number of variables in zt
Ingram mentions that one way to handle this difficulty is to "supplement the
stochastic elements of the model, either by the addition of measurement error to
the capital or consumption series or by the inclusion of additional exogenous
shocks," thereby differentiating between the actual, z" and latent, z~, random
variables. Definitionally, z, =z~+ Vr, where V, represents some "error in variable."
Substituting this out in equation (1) produces a relation in observables:
Z, = AZ,_1 + £, + v, - A vt-I. (2)
Ingram then outlines how MLE can be applied to this equation.
The main objection to this procedure is that some assumption needs to be made
about the nature of V,. Mostly, this is that v, is normally and independently distrib-
uted, but that would seem to be an heroic assumption for many of the SGE models
Ingram mentions. These are extremely stylized and almost certainly exclude vari-
ables that are important to actual realizations. In particular, this feature makes it
likely that v, will be correlated with Z~_I (and hence Z,_I)' thereby invalidating
standard estimation and inference measures. As an example of this phenomenon,
take Ingram's addition of a stochastic depreciation shock D, to the system-say,
D, = D+ £15,. Suppose that the modeler treats the depreciation rate as fixed at Dbut
agents correctly set it to the random value when optimizing. This means we would
have
COMMENTARY ON CHAPTER 2 51
= C~ - (1 - f3)eli,(k~_1 + Vk,_I)'
where c~ and k ~ are the solutions from the model setting the depreciation rate to
c5-that is, eli, = 0 and vk, = k,- k~ = -f3eli,(k~_1 + Vkl-l)' Hence, Ve , = C, - c~ =
-(1 - /3)eli,(k~_l + Vk,_I)' and this will be correlated with k~_I'
In such circumstances doubt is cast on the simulation methods for estimation
and inference outlined in Ingram's paper, and as practiced for example by Smith
(1993), as these simulate z~from (1) and then compare the implied moments with
those of the data, z,. The problem is that a wedge exists between these two vari-
ables, potentially creating a wide variety of relations between the two sets of
moments depending on the nature of V" a point made in Ingram, Kocherlakota,
and Savin (1993), Kim and Pagan (1994), and Watson (1993). In particular, it
needs to be emphasized that the popular strategy of comparing var(z,) with var(z~)
is meaningless unless accompanied by some statement about what is being as-
sumed regarding v,.
How do we deal with this quandary? One way to pose the question is to ask
how one should perform estimation and evaluation in the presence of specification
error of unknown form. Watson (1993) advanced the suggestion that bounds on
the moments for z, predicted by (1) be found that allow for v, and z~ to have
maximal correlation. Of course, this procedure produces only a measure of fit akin
to an R 2 and does not provide any information on how likely it is that the upper
bound will be encountered. For the latter one needs to have some idea of the data-
generating process (DGP) for Zr To gain this, one might fit a VAR to the data-
effectively making the assumption that v, can be represented as a linear function
of the past history of variables entering the VAR-and then treat that as the DGP.
As theoretical work on estimation and inference in the presence of specification
error shows, knowledge of the DGP leads to the ability to correct for the impact
of specification errors. Practically, the procedure requires that data be simulated
from the estimated VAR, and the properties of whatever estimators have been
used are then constructed from the empirical density function found by applying
the estimator to the generated data.
Finally, I want to address the question of the relation between the type of model
building described by Ingram and what has been the revealed preference of those
engaged in the practical application of macroeconometric models. One sometimes
has the impression that many academics in this area perceive that the problem is
simply one of getting the right parameters into the model. However, the models
Ingram describes are best thought of as useful for engaging in theoretical exer-
cises, specifically those that concentrate on examining what would happen within
an hypothetical economy as various shocks are administered to it. Their concern
is with the "big picture," and they rarely produce an understanding of the dynamics
52 MACROECONO~CS
of actual economies, a fact clearly seen in those instances where their predictions
regarding dynamics are matched up with the VAR's describing an actual economy,
such as Canova, Finn, and Pagan (1994). Nevertheless, there are certain features
of these models that make them attractive, and the following four are of particular
importance.
dealt with by Ingram in ways that are hard to describe in my limited space but that
stem from the fact that they aim to replicate actual economies. Examples might be
that agents are not endowed with full infonnation, and there is concern about
allowing for inertia or costs of adjustment when decisions are made. Rational
expectations are always found in some markets, such as those for financial assets,
but perhaps not in the short run for labor markets. Total factor productivity is
viewed as deterministic rather than stochastic, although they could almost cer-
tainly be moved in that direction if it was thought desirable. Choices on these
matters are model specific rather than universal. Because these models are in use,
it is important that academics be involved in scrutinizing them, and it seems timely
to draw attention to them and not to limit macroeconometrics to the type of models
set out by Ingram.
Notes
I. It is assumed by Ingram that one can write the system as in her (2.9}-that is, the matrix attached
c/,
to the contemporaneous variables {t, and A/ is nonsingular. Preston and Pagan (1982, pp. 298-308)
discuss rational expectations models for which this assumption fails and outline a "shuffle" algorithm
that may be applied to transform the system into one in which the matrix is nonsingular.
2. I would like to thank Beth Ingram for providing her GAUSS program for this task.
References
Bergstrom, A.R, and C.R. Wymer. (1976). "A Model of Dis-equilibrium Neo-classical
Growth and Its Application to the United Kingdom." In A.R Bergstrom (ed.), Statistical
Inference in Continuous Time Econometric Models. Amsterdam: North Holland.
Canova, F., M. Finn, and A.R Pagan. (1994). "Evaluating a Real Business Cycle Model."
In C. Hargreaves (ed.), Nonstationary Time Series Analyses and Co-integration. Ox-
ford: Oxford University Press.
Dixon, P.B., B.R Parmenter, A.A. Powell, PJ. Wilcoxen, and K.R Pearson. (1991). Notes
and Problems in Applied General Equilibrium Economics. Amsterdam: North Holland.
Dixon, P.B., B.R. Parmenter, J .M. Sutton, and D.P. Vincent. (1982). ORANI: A Multisectoral
Model of the Australian Economy. Amsterdam: North Holland.
Dotsey, M., and C.S. Mao. (1991). "How Well Do Linear Approximation Methods Work?
Results for Suboptimal Dynamic Equilibria." Journal of Monetary Economics 29, 25-
58.
Hall, RE. (1978). "Stochastic Implications of the Life-Cycle Permanent Income Hypo-
thesis: Theory and Evidence." Journal of Political Economy 86, 971-987.
Hansen, L.P. (1982). "Large Sample Properties of Generalized Method of Moments Esti-
mators." Econometrica 50, 1029-1054.
54 MACROECONOMETItlCS
Ingram, B., N. Kocherlakota, and N.E. Savin. (1993). "Explaining Business Cycles: A
Multiple Shock Approach." Mimeo, University of Iowa.
Johansen, L. (1960). A Multi-sectoral Study of Economic Growth. Amsterdam: North
Holland.
Jonson, P.D., E.R. Moses, and C.R. Wymer. (1977). "The RBA Model of the Australian
Economy." In Conference in Applied Economic Research. Reserve Bank of Australia.
Kim, K., and A.R. Pagan. (1995). "The Econometric Analysis of Calibrated Macroeconomic
Models." In M.H. Pesaran and M.R. Wickens (eds.), Handbook ofApplied Econometrics.
Blackwell.
King, R.G., C. Plosser, and S. Rebelo. (1988). "Production, Growth and Business Cycles
II: New Directions." Journal of Monetary Economics 21, 309-341.
Kydland, F., and E.C. Prescott. (1991). "The Econometrics of the General Equilibrium
Approach to Business Cycles." Scandanavian Journal of Economics 93, 161-178.
Laxton, D., and R. Tetlow. (1993). "Government Debt in an Open Economy." Bank of
Canada Discussion Paper.
Mandlebrot, B. (1963). 'The Variation of Certain Speculative Prices." Journal of Business
36,394-419.
Mao, C.S. (1990). "Hypothesis Testing and Finite Sample Properties of Generalized Method
of Moments Estimators: A Monte Carlo Study." Mimeo, Federal Reserve Bank of
Richmond.
Masson, P., Symansky, R. Haas, and M. Dooley. (1988). "Multimod: A Multi-region
Econometric Model Part II." IMF Staff Studies for the World Economic Outlook.
McKibbin, WJ., and J. Sachs. (1991). Global Linkages: Macroeconomic Interdependence
and Co-operation in the World Economy. Washington, DC: Brookings Institution.
Murphy, C.W. (1988). "An Overview of the Murphy Model." In M.E. Bums and C.W.
Murphy (OOs.), Macroeconomic Modelling in Australia, supplementary conference is-
sue of Australian Economic Papers, 61-68.
Pagan, A.R., and Y. Jung. (1993). "Understanding the Failure of Some Instrumental Vari-
able Estimators." Mimeo, Australian National University.
Pagan, A.R., and G.W. Schwert. (1990), "Testing for Covariance Stationarity in Stock
Market Data." Economics Letters 33, 165-170.
Phillips, P.C.B. (1989). "Partially Identified Models." Econometric Theory 5, 181-240.
Preston, AJ., and A.R. Pagan. (1982). The Theory of Economic Policy. Cambridge: Cam-
bridge University Press.
Rose, D.E., and lG. Selody. (1985). "The Structure of the Small Annual Model." Technical
Report No. 40, Bank of Canada, Ottawa.
Smith, A.A. (1993). "Estimating Non-linear Time Series Models Using Simulated Vector
Autoregressions." Journal of Applied Econometrics 8, S63-S84.
Staiger, D., and J.H. Stock. (1993). "Instrumental Variables Regression with Weak In-
struments." Mimeo, Kennedy School of Government, Harvard University.
Taplin, B., P. Jilek, L. Antioch, A. Johnson, P. Parameswaran, and C. Louis. (1993).
Documentation of the Treasury Macroeconomic (TRYM) Model of the Australian
Economy. TRYM Paper No.2, Australian Treasury. Canberra.
Watson, M.W. (1993). "Measures of Fit for Calibrated Models." Journal of Political
Economy 101, 1011-1041.
COMMENTARY ON CHAPTER 2 55
Theraption the monk wouldfrequently interrupt his prayers to follow tenderly the
love-making games ofthe swallows, because that which isforbidden to the nymphs
is allowed to the swallows.
-Marguerite Yourcenar
Introduction
57
58 MACROECONOMETIUCS
and, more recently, as a way to compare actual data with the time series generated
by artificial economies with calibrated parameters.
The purpose of this chapter is to describe the salient aspects of the original
VAR methodology and of its current extensions, discuss the validity of policy
analyses conducted with VAR models and their meaning, and compare the meth-
odology with other competing approaches for applied research. Because of this
choice we will only tangentially touch other matters that are of interest to VAR
practitioners (such as forecasting or simulation comparisons with theoretical
models). We will try to separate and distinguish as much as possible statistical
from economic issues in order to highlight the stages where crucial assumptions
are made. Therefore the next section begins by discussing questions concerning
the specification and estimation of VAR models. There we present the approach
followed by users of unrestricted VAR models and contrast it with the approach
followed by rational expectation econometricians. We also discuss the so-called
Bayesian VAR (BVAR) approach to specification and estimation and relate the
methodology to the index models literature.
Section 3 deals with inference and with issues of interpretation of the results.
In this section we provide a discussion of four important issues. First, we highlight
the major features of what we call semistructural approach to policy analysis and
contrast its philosophy with the one of rational expectation or generalized method
of moment (GMM) econometricians and with the one of calibrators. Second, we
discuss in detail the economic issues concerning the identification of a generic
model of economic behavior from unrestricted VAR models. Third, we argue that
the range of economic questions that are well posed within the context of the type
of generic models of economic behavior that can be recovered from the VAR is
sufficiently thick and does not necessarily violate the dictates of the Lucas's
critique. Finally, we explain how innovation accounting exercises conducted on
the generic structural model that is recovered via identification can be used to
study questions of interest to macroeconomists and policymakers.
The fourth section examines three problems that may arise with the inter-
pretation of the results obtained from semistructural VAR analyses. The first has
to do with the correct number of shocks buffeting the system. The second with the
recoverability of structural shocks. The third with time aggregation. A simple
example highlights some of these problems.
The final two sections briefly examine some issues concerning the character-
ization of data interdependencies and unconditional forecasting exercises and
conclude by stressing where critics of VAR models have pointed their guns and
suggesting a possible avenue for future developments.
VARMODELS 59
It is well established by now that either the exact or the approximate solution of
many stochastic dynamic-equilibrium models comes in the fonn of a restricted
VAR (see Hansen and Sargent, 1980; King, Plosser, and Rebelo, 1988). Many of
these models are in the neoclassical tradition, but examples of models that deliver
dynamic solutions in the fonn of VARs also appear in the recent new Keynesian
literature (see Taylor, 1980). One approach to the specification of a VAR, out-
lined, for example, in Hansen and Sargent (1980), is to tie down its parameterization
to the underlying theory, estimate the unknown parameters of the model via maxi-
mum likelihood techniques, and examine the validity of the restrictions imposed
by the theory using standard techniques such as likelihood ratio tests. Apart from
the issue of estimation of the unknown parameters, the idea of tying down the
parameters of the VAR to the theory and the testing apparatus developed by
Hansen and Sargent can also be applied to the problem of evaluating dynamic
economic models whose parameters are calibrated (see Canova, Finn, and Pagan,
1993).
Despite the fact that the VAR parameters are tied down to the "deep" para-
meters of the theory, there are at least two elements of arbitrariness that remain
in specifying the empirical VAR corresponding to the theory. First, economic
theory does not usually provide indications regarding the lag length of the VAR
model. One can take two approaches in this respect: either a priori choose a lag
length and verify that the results are independent of this auxiliary assumption or
let the data choose the correct lag length using some optimal statistical criteria
(such as Akaike, 1974, or Schwartz, 1978). Second, many dynamic models deliver
solutions for the vector of endogenous variables whose covariance matrix is sin-
gular because there is a larger number of endogenous variables than shocks, so in
order to undertake a meaningful estimation one must probabilistically complete
the model by adding other sources of disturbance. These new sources of distur-
bance may come, for example, from measurement errors. In general, it is assumed
that agents have an infonnation set that is larger than the one of the econometrician.
In this case the estimated VAR is nonsingular because important sources of
dynamics are omitted from the model. Within this approach the econometrician
must therefore choose which variables are either measured with error or not ob-
servable, and these choices must be verified and substantiated before the results
are discussed.
The specification of a VAR model, however, need not to be tied down to any
particular parameterization of the theory. Sims (1980a) suggested using an unre-
stricted VAR approach that is not linked to any particular theory. The major
building block of the unrestricted VAR methodology is the Wold theorem. Com-
plete technical details of the mathematics needed are contained in Rozanov (1967)
60 MACROECONOMETIUCS
or Quah (1992), while a simple and thorough presentation of the theory appears
in Sargent (1986, ch. 11) or Canova (1994).
The Wold theorem states that one can decompose any vector stochastic process
x, of dimension m at every point in time t into the sum of two orthogonal com-
ponents: one that is predictable based on the information available at time t - 1
and one that is unpredictable. Since the above decomposition holds at each t,
repeated application of this decomposition to the vector x, yields
where F, is the information set available at t, F_~ is the information set available
in the infinite past, e,_j is the set carrying the new information at each t - j, and
P, is an operator that geometrically projects x, onto various information sets. The
first component appearing on the right side of (3.1) is the detenninistic part of X,-
that is, the part that is predictable given information contained in the infinite past,
and the second component is the "regular" part, the part of x, carrying the new
information that becomes available at each t. While it is not necessary for the
theory to be applied, it is typical to restrict the projection to be linear, so that it
possible to write P'[x,1 F_~] =Hj,d,_j =H,(e)d" P,[x,1 e,_j] = Cj,e,_j = C,(e)e, 'Vj,
p
where H(e) = Ho + HIe + ... + Hp, C(e) = Co + cle + ... + cpe and e is the lag
operator, Hj , and Cj ' are matrices of coefficients, d,a set of detenninistic variables
and e,_j = xl-j - P,_j[x,_j I F,_j_l]. By construction the sequence {er-j};'o is a zero
mean process that is uncorrelated with all of its past-that is, e, represents the
"news" or the "innovations" in x, and therefore should be unpredictable using past
information. The coefficients Cj ' satisfy two conditions: (1) the elements of x, are
contemporaneously uncorrelated-that is, CjO = I, and (2) the sequence of x,
defined by (3.1) is bounded-that is, !:j:oC], < 00. If this condition did not hold,
news that occurred infinitely far in the past would have a nonnegligible effect on
the process at t.
It is also typical to assume that x, is covariance stationary-that is, the
autocovariance function of x, depends on the distance between two elements not
no their position in time (Ex,x;_s depends on s but not on t), so Hj , and Cj ' become
independent of t. This restriction is made purely for convenience and is not needed
for the theory to be applied. Because economic time series often do not satisfy this
assumption, transformations of the data are usually required to induce this property.
When the restrictions (1) and (2) on the Cj coefficients are satisfied there exists
an autoregressive representation that expresses e, as a linear combination of cur-
rent and past x,'s of the form
In choosing the variables that enter the VAR, one can appeal to economic
theory, empirical observations, or experience. In principle, one would like to in-
clude all variables that are likely to have important interdependencies. In practice,
small-scale models are used because of the limited size of existing data sets.
Implicit in the choice of variables is therefore a process of marginalization-that
is, the joint probability density of the selected VAR has been marginalized with
respect to all other potentially relevant variables (see Clements and Mizon, 1991).
This assumption has raised some controversy. Some (see McNees, 1986; Fair,
1988; Kirchgassner, 1991) consider it dubious and a potential source of incorrect
inference in small-scale VAR models. Others considers it innocuous as long as the
statistical model is well specified.
In general, the polynomial r( f) appearing in equation (3.2) will have an infinite
number of elements for any reasonable specification of the polynomial C(f). That
is to say, under the regularity conditions we have outlined, an unrestricted VAR
will approximate arbitrarily well any vector stochastic process X, when the lag
length of the model gets arbitrarily large. However, because of data limitations,
one is forced to use a finite-order VAR model in the analysis. In this case, one can
either consider the selected VAR as the true data-generating process (DGP) for the
actual data, as a finite-order approximation to an underlying infinite-order linear
model, or as a first-order Taylor approximation of a (in)finite order nonlinear
model. Which of these three approaches is taken is essentially irrelevant in prac-
tice for two reasons. First, nonlinearities are typically of marginal interest in the
VAR context, since monthly and quarterly macroeconomic variables are usually
well approximated by linear time series processes (see Brock and Sayers, 1988).
Second, except in the case of the computation of asymptotic standard errors and
in detennining the finite sample properties of estimators of functions of the VAR
parameters, it will not matter which of the first two points of view is taken (see
Lutkepohl, 1991).
62 MACROECONOMETRICS
In choosing the lag length of a finite-order unrestricted VAR model one must
weigh two opposing considerations: the "curse of dimensionality" and the correct
specification of the model. Since the number of parameters to be estimated quickly
increases with the number of lags of the system, systems of moderate size become
highly overparametrized relative to the total number of observations. This leads to
poor and inefficient estimates of the short-run features of the data. A lag length
that is too short, on the other hand, leaves unexplained information in the distur-
bance term, generating a statistical model where only a subset of the existing
information is used to characterize the data and inducing spurious significance in
the coefficients. The tradeoff between overparameteization and oversimplification
is at the heart of almost all statistical selection criteria designed to choose the lag
length of the autoregression (see Akaike's AIC, 1974; Schwarz's, 1978 SIC
criteria; the modified likelihood ratio (MLR) procedure of Sims, 1980a; or the
selection criteria suggested by Lutkepohl, 1991). While all these procedures are
asymptotically equivalent, their small sample properties differ significantly and
may depend on the true lag structure ofthe data generating process (see Nickelsburg,
1985; Lutkepohl, 1986).
One advantage of choosing the lag length of the VAR according to one of the
above criteria is that the resulting model can be estimated using very simple tools.
Because only predetermined variables and deterministic functions of time appear
as right-side variables, and because the disturbances e, are serially uncorrelated,
OLS estimates of the VAR coefficients are consistent. In addition, because of the
unrestricted nature of the model, single equation methods are (asymptotically)
efficient. This is because a VAR model is a dynamic seemingly unrelated model
(see Judge et al., 1980). Such a model has the property that the same right-side
variables appear in each equation. Under this condition Zellner (1962) showed that
single-equation methods are as efficient as systemwide methods no matter what
the contemporaneous correlations among the disturbances are. In addition, when
the e's are normal, OLS is efficient.
One drawback of choosing the lag length of the model using these selection
procedures is that all the variables entering the system have identical lag lengths,
which need not be a valid economic or statistical restriction, may reduce the
efficiency of the estimates and, in some cases, even bias the results. To overcome
this problem Hsiao (1981) suggested an approach that starts from univariate auto-
regression and sequentially adds lags and variables using predictive criteria. Ac-
cording to this criteria an additional variable should enter a univariate autoregression
with a certain number of lags if it helps to improve the forecast of the endogenous
variables-that is, if it Granger causes the left-side variable. Two drawbacks of
this approach should also be mentioned. First, since the results of the predictive
tests are treated as exact, the successive examination of hypothesis with economic
content may be sensitive to type I errors committed at the specification stage.
VAR MODELS 63
Second, since the same variables do not appear in all equations, systemwide
methods (such as full-information maximum likelihood (FIML» are needed for
efficient estimation. An alternative procedure to selection of the length of the
VAR, which goes in the opposite direction of Hsiao's specific-to-general ap-
proach, has been popularized in the works of Hendry and Mizon (1990) and
Clements and Mizon (1991), and it is thoroughly discussed in Chapter 4 of this
book.
The idea that macroeconomic time series are stationary around, say, a deter-
ministic polynomial function of time has come under debate in the last decade (see
the original study by Nelson and Plosser, 1982). However, recent work by Perron
(1989) and Dejong and Whiteman (1991) suggests that the evidence in favor of
unit roots is not so overwhelming. While covariance stationarity is not needed to
derive the Wold theorem, any estimation and testing procedure requires time
invariance of the parameters of the model. Therefore, in specifying an unrestricted
VAR model, a researcher has to take a stance on the sources of nonstationarity
(deterministic or stochastic) and on the possible presence of shifts that may occur
from time to time in economic time series. The choice made has implications for
estimation and testing. For example, Phillips (1988) shows that knowledge of the
order of integration of the variables of a regression and of the nature of determin-
istic components can be important in designing optimal inference. This is because
testing hypotheses on coefficients of integrated variables requires nonstandard
asymptotic theory. These considerations suggest that the appropriate modeling
strategy to determine the long-run properties of each variable of the VAR model
should be based on univariate unit roots tests (such as Dickey-Fuller, 1979; Phillips
and Perron, 1986; Stock and Watson, 1989). However, as Engle and Granger
(1987) have indicated, univariate unit root procedures may generate too many unit
roots in a VAR. The solutions proposed to deal with this problem differ. Engle and
Granger suggest writing the model in a vector error-correction form and use a two-
step least-square procedure to estimate the parameters of interest. Johansen (1988)
and Johansen and Joselius (1990) suggest jointly estimating cointegrating vectors,
if they exist, and the coefficients of the VAR model by full-information maximum
likelihood under rank restrictions on the matrix of long-run coefficients. Sims,
Stock, and Watson (1990) suggest using an unrestricted VAR in levels and check-
ing the order of integration of the variables and the number of cointegrating
restrictions in order to know which asymptotic distribution applies.
To summarize, the classical approach to specification and estimation of VAR
models involves two steps. First, one chooses a model that is dynamically well
specified (in terms of variables included, lag length, noncorrelation, and, possibly,
normality of the residuals), extracts as much information as possible from the data
and tests for the order of integration of the data, taking into account the possibility
of regime shifts, segmented trends, and so forth. Second, one can either transform
64 MACROECONOMETRICS
the system and estimate the VAR coefficients using the two-step procedure of
Engle and Granger, estimate the original VAR model under rank restrictions using
the maximum-likelihood approach of Johansen, or estimate the coefficients of the
untransformed, unrestricted model by OLS, equation by equation. In the first two
cases testing of hypotheses on the coefficients of the transformed system can be
undertaken using standard asymptotic theory. In the third case, the results of Sims,
Stock, and Watson must be applied to test hypotheses of interest. As an interme-
diate step, it is possible to reduce the dimensionality of the system by eliminating
either variables or lags that appear to be unimportant in explaining variations of
the endogenous variables in one or more equations while maintaining a dynami-
cally well-specified system. If this intermediate step is taken, single-equation
methods of estimation of the second-step regression in Engle and Granger's pro-
cedure or of the untransformed system are no longer efficient.
An alternative strain of literature, originating from Litterman (1980, 1986) and
described in Todd (1984), takes a Bayesian perspective to the specification of
unrestricted VAR systems. The starting point of the approach is that in specifying
econometric models one attempts to filter as much information as possible from
the data. However, instead of relying on classical hypothesis testing or economic
theory to decide whether a particular variable or lag should enter the VAR, these
authors use a symmetric "atheoretical" prior on all variables to trade off over-
parametrization with oversimplification. The reason for taking an alternative route
to the specification problem is that there is a very low signal-to-noise ratio in
economic data and economic theory leaves a great deal of uncertainty concerning
which economic structures are useful for inference and forecasting. Because highly
parametrized unrestricted VAR models include many variables in each equation,
the extraction filter is too wide and the noise obscures the relatively weak signal
present in the data. The prior these authors employ is characterized by a small
number of parameters and acts as an orientable antenna that, when appropriately
directed, may clarify the signal. A specification search a-I a Leamer (1978) over the
parameters of the prior tunes up the filter for an optimal extraction of the infor-
mation contained in a given data set. This approach to specification has two
advantages over classical specification approaches. First, it avoids strong restric-
tions on which lags should enter the autoregression and refrains from specification
testing procedures that may bias the examination of economic hypotheses. Second,
it is flexible enough to allow different lag lengths to enter in different equations.
The prior typically appended to VAR models is nonstandard from the point of
view of Bayesian analysis; it is objective in the sense that it is based on experience
and reflects accepted rules in the forecasting community and is, in most cases,
atheoretical because the restrictions implied by the prior have no economic inter-
pretation. There are three basic observations underlying the specification of the
prior. First, univariate random walks well represent the features of many series.
VARMODELS 65
Second, the information contained in the far past of the series, has, in general, little
importance in predicting current movements in the series. Third, both the condi-
tional and the unconditional distribution of many variables seem to be shifting
over time.
Because an unrestricted Bayesian VAR (BVAR) is nothing more than a state
space-time-varying coefficient model, techniques developed for that framework
can be used to estimate the unknown parameters of the model. In particular,
maximum-likelihood estimates of the parameters of the model can be obtained by
passing through the sample recursively with the Kalman filter algorithm (see
Doan, Litterman, and Sims, 1984). In addition, in the special case considered by
Litterman (1986)-a BVAR model is simply a VAR with a dummy observation
appended to the system. In this special case the estimation of the model parameters
can be carried out with a mixed-type estimation (see Theil, 1971). The result is a
restricted estimator that shrinks the data toward the information contained in the
prior.
Contrary to the classical approach previously discussed, the presence of trending
variables in the system does not cause particular problems to this or any generic
Bayesian approach (see Sims, 1988b; Sims and UIhig, 1991). Inference in BVAR
models is based, in fact, on the likelihood principle. The approach requires normal
residuals, which may not always be a reasonable assumption, and good priors,
which are not always easy to obtain, but is entirely invariant to the size of the
dominant root of the system. As a consequence, no pretesting is need to determine
which asymptotic distribution applies (the standard or the unit root one), and
inferential procedures are considerably simplified.
It should be finally noted that a BVAR model is flexible enough to include as
special cases several models that have been used to characterize the properties of
macro and financial time series. Apart from the direct links with time varying
coefficient (TVC) models (see Nicholls and Quinn, 1982), ARCH (Engle, 1982),
ARCH-M (Engle, LiIien, and Robbins, 1987) models, bilinear models (Granger
and Anderson, 1978), discrete regime shift models (Hamilton, 1989), and subor-
dinated models (see Gallant, Hsieh, and Tauchen, 1991) can be nested and tested
within a general BVAR specification (see Canova, 1993).
A final approach to specifying VAR' s is linked to the observable index model
literature (see Sargent and Sims, 1977; Quah and Sargent, 1993). The idea is to
start from a completely unrestricted VAR model and examine whether an m-
dimensional process XI can be explained or predicted by past values of an index
series 711 of dimension r < m. The economic rationale behind the scheme is that
there is a low dimensional vector of variables (the states of the economy) that is
responsible for most of the dynamic interdependencies of the economy. With
these lenses, observable index models can be simply seen as providing an alter-
native focus to the overparametrization/oversimplification tradeoff. But there is
66 MACROECONOMETRICS
also an alternative interpretation that may shed some light on the links between
this approach and others. If we assume that one of the indices is a nonstationary
process that is common to all series, this approach amounts to specifying and
estimating a vector error-correction model or a common trend model (as in Stock
and Watson, 1988) so that all the tools developed for these frameworks apply here
as well. For the case in which all indices are stationary, Reinsel (1983) describes
selection procedures for these models together with the theory for identification of
the parameters, their estimation by maximum likelihood, and one approach to test
the cross-equation restrictions implied by the index structure.
The previous section dealt with the specification and the estimation of VAR models.
While these are important issues from a statistical point of view, and model selec-
tion is of crucial importance for forecasting purposes, the specification and esti-
mation of an unrestricted VAR are only preliminary steps in undertaking economic
analyses of the data. Because an unrestricted VAR is a reduced-form model, it is
uninterpretable without reference to theoretical economic structures even when it
is correctly specified and efficiently estimated. In other words, economic theory
cannot simply be kept in the background during the analysis and used only as an
organizing principle for the empirical research, as, for example, is the case when
producing "stylized facts" in the empirical real business cycle literature. Instead,
it must be employed (either in a weak or strong sense) in order to extract economi-
cally relevant information from the reduced form VAR innovations and moving
average coefficients. This does not necessarily mean that one must have a well-
specified economic model in mind to undertake the analysis (as is the case, for
example, for Hansen and Sargent, 1980, 1991). In order to extract information
about the behavior of agents in the economy one can appeal to a class of models
that impose very generic restrictions on the data (such as long-run neutrality or
superneutrality) or simply use informational constraints that are linked to the flow
of information in the economy (see also Quah, 1991).
The issue of providing an economic interpretation of unrestricted VAR results
is therefore linked to the question of the identification of a structural model from
reduced form estimates. The terms structure and structural have many meanings
in the economic literature. Some economists label structural a model where all
parameters have economic interpretations in terms of preferences and technolo-
gies (Hansen and Sargent, 1980). Others claim that a model is structural if the
consequence of a specific class of actions can be predicted by modifying part of
VARMODELS 67
the model (see Koopmans, 1963), if it is invariant to a specific class of actions (see
Sargent and Lucas, 1980), or if it refers to the use of economic a priori hypothesis
(Hendry and Mizon, 1990). Finally, some claim that a model is structural if it is
in a particular mathematical form (see Harvey, 1985). Within the VAR literature,
it is typical to term a model structural if it is possible to give distinct behavioral
interpretations to the stochastic disturbances of the model. An example here may
clarify our terminology. Consider a bivariate VAR with output and money. Then
a structural model is obtained when it is economically meaningful to name the
disturbances of the model-say, aggregate demand shocks, money supply shocks,
and so on. From the perspective of identification, VARs are not dramatically
different from those macroeconometric models that Sims criticized in his original
(1980a) article. There is, however, a subtle difference in the approaches that con-
cerns the style of identification employed. Traditional macroeconometric models
typically were of large scale and attempted to identify and recover a structural
model making frequent and extensive use of exclusion restrictions on the lags of
some variables. Users of this type of models had in the back of their minds the
dictates of the Cowles Commission that variables can be classified into endo-
genous and exogenous, where exogenous were those variables that were deter-
mined outside of the system and could be manipulated independently of the
remaining variables. Exclusion restrictions on the lags of some variables were the
practical tools to achieve this distinction.
Sims, echoing Liu (1960), criticized this procedure on the basis that there was
no economic justification for the massive amount of exclusion restrictions em-
ployed to identify models and that these restrictions were routinely imposed with-
out too much attention to the underlying economic structure. Given that economic
theory placed very weak restrictions on the reduced-form coefficients and was
essentially silent in deciding which variables must enter a reduced-form model,
empirical research ought to be based on small-scale models that required a small
number of constraints to be interpreted. In addition, and more important, users of
unrestricted VAR models typically refrain from imposing restrictions on lag struc-
ture and, instead, attempt identification via restrictions on the contemporaneous or
the long-run effects of behavioral shocks or on both. One reason for this choice
is that the classification of variables in exogenous and endogenous is meaningless
from the point of view of modem dynamic general equilibrium theory. In addition,
theory has typically little to say about the time response of the variables of the
system to shocks, while it has somewhat stronger predictions on lag time needed
for the response to take place-for example, theory may have something to say on
how monetary policy shocks affect the real side of the economy or on their long-
run effects but not much to say regarding the magnitude of the responses three
periods after a monetary shock has occurred. Interesting restrictions on the con-
temporaneous or the long-run impact of behavioral shocks are not numerous, but
68 MACROECONOMETIUCS
they are typically sufficient to just identify the disturbances of a small system of
variables.
It is useful at this point to stress that the scope of economic analysis undertaken
with VARs is not philosophically different from the analyses undertaken under the
rational expectations econometric program pioneered by Hansen and Sargent (1980,
1991). Also, in this case, the objects of interest are various sources of innovations
to an agent's infonnation set and their effect on the endogenous variables of
the system. However, the way these objects are recovered from the reduced-fonn
VARs is different. Rational expectations models typically involve an extensive set
of cross-equation restrictions that, when imposed on the VAR, allow the recovery
of the so-called deep parameters of agents' preferences, technologies, and con-
straints. Because of the interrelationship between forcing variables and variables
determined in equilibrium by agents' behavior, there are typically more con-
straints than reduced-fonn parameters, so a rational expectation econometrician
can give additional content to the theory by testing the validity of these additional
restrictions.
A VAR econometrician can be thought of as a rational expectation econo-
metrician who is skeptical of many of the restrictions that a particular fonnula-
tion of dynamic economic theory imposes on the VAR coefficients. Therefore, in
order to produce a structural interpretation of the VAR model, he uses only a
limited number of these constraints and "lets the data speak" regarding certain
policy questions. A rational expectation econometrician, on the other hand, uses
all the restrictions that a particular fonnulation of the theory imposes in order to
produce refutable structures from the reduced-fonn VAR model. The main differ-
ence here is therefore on the extent of the economic restrictions employed, not on
their merit or their usefulness. To put the argument in another way, the number
of restrictions needed to identify behavioral disturbances is substantially smaller
than the number of restrictions that a rational expectation econometrician imposes
on the VAR to identify structural parameters and validate the model. It is also clear
that, because a VAR econometrician uses only a subset of the theoretical restric-
tions, he can undertake only a limited set of interesting economic experiments. For
example, he cannot undertake experiments that change the probability distribu-
tions of the behavioral disturbances. In this case, the famous Lucas's critique
applies, and the exercise are meaningless. However, Sims (1982, 1988a) claims
that although theoretically appropriate, the Lucas's critique should be nothing
more that a cautionary footnote for skilled users of unrestricted VAR models.
There are two reasons for this. First, the nonnal course of policy actions does not
involve changes in the probability distribution of policy disturbances but instead
can be seen as drawing different innovations from the existing distribution of dis-
turbances. Second, forecasting the results of regime changes that have never been
experimented with in the past or of situations that involve dramatic changes in the
VARMODELS 69
difference is that calibrators directly estimate the deep parameters of the model
while a rational-expectation econometrician estimates reduced-fonn coefficients
subject to the identifying restrictions imposed by the theory. One additional dif-
ference should also be noted. Calibrators estimate parameters imposing economic
restrictions on the error of the model (for example, they assume that the model's
parameters are chosen so that it exactly fits the steady-state properties of the actual
economies) while a RE econometrician estimates parameters imposing statistical
assumptions on the error of the model (for example, nonnality).
Identification
The starting point of the identification process is the assumption that there is a
class of dynamic theoretical models that deliver a solution for the endogenous
variables of the model of the fonn
where f is the lag operator, the matrix polynomial B( f) has all its roots not inside
*
the unit circle, Bo I and where E(utu~) is a full-rank matrix and where the
elements of B(l) = ~Bj may be, in principle, different from zero. The u's are
primitive exogenous forces not directly observed by the econometrician that buffet
the system and cause oscillation. Because the u's are among the primitives of the
model, they are typically assumed to be contemporaneously and serially uncor-
related and can be given fairly specific behavioral interpretation. Models of this
type can be found in Keynesian macro literature (see Taylor, 1980; Blanchard,
1989), in the rational-expectations literature (see Quah, 1990; Hansen, Roberds,
and Sargent, 1991) and in the real business cycle literature (see King, Plosser, and
Rebelo, 1988).
We assume that an econometrician investigates the properties of an unrestricted
VAR model of the fonn
The moving average representation for the theoretical model and the data are,
respectively,
y, = M(f)d, + N(f)Bo1Su,
= M(f)d, + N(1)Bo'Su, + N*(f)Bo1S!!.u, (3.5)
y, = H(f)d, + C(f)e,
=
where y, is a stationary inducing transformation of x" such as y, (I - Ox" !!.u,
= u, - U'_I' !!.e,= e, - er-l' M(f) = F * iJ (fr l , H(f) = G * t-I(f), N(f) =iJ-1(f),
C(f) = t-'(f), and where N*(f) :; N(f) - N(l) and C*(f):; C(f) - C(1).
I-f I-f
Matching the coefficients of the moving average representations (3.5) and (3.6),
we have
The question of identification can then be posed in the following way: under
what conditions is knowledge about the reduced-form VAR parameters and inno-
vations sufficient to recover objects that may be of interest to the economic analyst?
Clearly, unless the restrictions imposed by modem dynamic macroeconomic models
are not taken into account, there is little hope of disentangling the "deep" para-
meters of preferences and technologies from the VAR coefficients and, as a con-
sequence, little hope of recovering a model which is structural according to Lucas's
terminology. Sims (1986), however, has argued that there is no need to recover the
deep parameters of preferences and technologies to predict the consequences of
specific policy actions. In other words, with meaningful economic restrictions it
is possible to extract from a VAR a generic structure, which may not satisfy
Lucas's criteria for policy intervention but is still useful to indicate the impact of
policy actions that fall within the realm of historical experience. And this may be
everything that an econometrician is interested and entitled to do.
To distinguish the analyses that originated from Sims's ideas from the structural
72 MACROECONO~CS
Practical Details
For many models of interest, the relationship between a particular structure and
the reduced-form VAR model holds at all lags. Therefore, one is free to choose
restrictions for contemporaneous correlations, for long-run correlations, or for
both. Initial users of VAR models (see Sims, 1980b, 1982) were not concerned
with nonstationarities. Therefore the long-run multipliers C(l) and N(l) were
assumed to be zero, and the restrictions used to identify the system were of
the form (3.7) and (3.8). Within this class of restrictions, a typical choice was
to assume that the unobservable behavioral disturbances contemporaneously
impacted on one equation only-that is, S = I-and that the observables had a
particular contemporaneous structure-that is, 8 0 lower triangular (see Sargent,
1979; Sims, 1980b; Litterman and Weiss, 1985). This choice was originally
proposed as a mechanical instrument to avoid data mining and to reduce the
74 MACROECONOMETRICS
the true behavioral shocks driving the system and as such uninterpretable. As we
observed, dynamic structural economic models may place restrictions on the cor-
relation among observables (the Bomatrix), on the correlation among unobservables
(the S matrix), or on both. In choosing one ofthese structures one ends up selecting
a particular model for the correlation among variables. As emphasized by Blanchard
(1989), absent meaningful economic motivation, such a choice is arbitrary.
The answer to this important objection has been to impose more economic
structure on the restrictions employed to identify the system. The restrictions
however, still remain generic, in the sense that they are generated by a large class
of models that cover most of the existing camps. Four major approaches have been
proposed. The first one, which emerges in the work of Bemanke (1986), Blanchard
and Watson (1986), Blanchard (1989), Evans (1989), and Gali and Hammour
(1991), maintains the basic identification style of earlier works but places restric-
tions on both the Boand S matrices and provides explicit economic justification for
each choice. The second one, contained in the work of Sims (1986) and Canova
(1991), instead of employing both debatable and scarce economic restrictions,
uses knowledge about the flow of information in the economy to restrict the
matrices B o and S and to recover generic semistructural models. The third one,
pioneered by Blanchard and Quah (1989), Shapiro and Watson (1989), and King
et al. (1991) exploits the nonstationarity of the data to impose restrictions on the
long-run multipliers of the structural model-that is, the N(I) matrix. A fourth
approach, suggested by Gali (1992), mixes aspects of the first and the third
approaches to achieve identification.
The first approach is sufficiently self-explanatory and does not require further
comment. Because identification achieved via the second approach does not typi-
cally involve restrictions derived from economic theory, a simple example may
clarify the usefulness and the limits of this approach. Suppose two of the variables
of the system are nominal GNP and money and that a VAR econometrician is
interested in studying whether monetary policy has effects on nominal GNP. Be-
cause of lags in data collection, it is known that GNP data becomes available only
once every three months. Given this informational delay, it is reasonable to sup-
pose that within a month, GNP is not a variable in the reaction function of the
monetary authorities. This informational lag therefore provides identifying infor-
mation that can be used to recover a behavioral rule used by the monetary author-
ity. It is important to stress that the above identification procedure does not imply
that over longer horizons (even a quarter) the monetary authorities do not use
information contained in GNP data in setting targets for the growth rate of the
money supply. There are two advantages of this approach. First, because this ap-
proach is less theoretically bound, it can be more useful in investigating contro-
versial relationships that may otherwise require stringent assumptions on the nature
of the behavioral shocks impinging on the system. Second, because informational
76 MACROECONOMETIUCS
delays are very common, there are often more than enough restrictions to identify
particular shocks or, in some cases, the entire vector of behavioral shocks.
The use of long-run restrictions to identify a VAR has been motivated as a way
of generating an unrestricted "pennanent-transitory" decomposition for time
series with unit roots that overcomes the lack of interpretability of traditional de-
compositions such as those of Beveridge and Nelson (1981) or Watson (1986). A
VAR identified with long-run restrictions imposes an arbitrary but specific struc-
ture on the pennanent and transitory components of the series and provides a
useful measure of the relative importance of pennanent components in explaining
cyclical fluctuations in the system. However, while the use of long-run restrictions
to identify the system has appealing features from the point of view of economic
theory, the number of meaningful and uncontroversial long-run restrictions is
limited, so for systems of medium size one hopes to have, at best, enough con-
straints to be able to just identify the model. In addition, disturbances identified
via long-run restrictions cannot be in general related with interpretable economic
news to agents infonnation set. For example, a shock that has long-run effects can
be classified as a supply disturbance (such as a technology shock) or a demand
disturbance (a fiscal shock). Alternatively, a shock that has only transitory impact
could be either real or monetary. In other words, disturbances recovered via long-
run restrictions are indices comprising linear combinations of "economically
meaningful disturbances" with no immediate behavioral interpretation.
As already mentioned, when the variables of the VAR are nonstationary but
cointegrated, the burden of finding identifying restrictions is lighter, since the pre-
sence of cointegration implies that the number of shocks driving the system (the
so-called common trends) is of smaller dimension than the vector X" In practical
tenns, the fact that the variables of the systems are nonstationary but cointegrated
means that N( I) is a matrix of reduced rank. If the dimension of X, is m and there
are p common trends, then the rank of N(l) is m - p. This means that there are only
m - p independent sources of disturbances that have long-run effects. Hence, if
cointegration is present, we need only m x m - p restrictions instead of m x m to
be able to identify all the behavioral disturbances of the model. King et al. (1991),
Quah (1991), and Wickens (1992) examine this situation in detail. Two points
regarding the usefulness of cointegration restrictions to identification can be made.
First, because unit root and cointegration tests appear to have very little power
against the trend stationary alternative, the results may be very sensitive to the
treatment of trends, and identification schemes that condition the economic analy-
sis on weak statistical findings should be taken with great care. Second, although
the imposition of cointegration restrictions helps identification by decreasing the
number of restrictions needed to undertake a semistructural analysis, it helps, in
a somewhat uninteresting way, by shifting the burden of finding an underlying
structure away from economic theory onto the characterization of the statistical
VARMODELS 77
properties of economic time series. In other words, the relevant question becomes
one about the existence of unit roots and cointegration rather than of economic
behavior.
If the statistical conditions for identification are satisfied and the restrictions are
placed either on Bo or on S or N(1), one can employ either an indirect least-squares
(ILS) procedure as in Sims (1986) or a two-stage least-squares procedure (2SLS)
as in Blanchard (1989) to estimate the free parameters of the model. The first
procedures involves first the computation of unrestricted VAR coefficients and of
the variance covariance matrix of the VAR shocks. Then, given a particular the-
oretical structure, one manipulates these quantities to obtain estimates of the free
parameters of the model. A 2SLS procedure imposes the identifying restrictions
directly on the VAR system and obtains estimates of the free parameters using an
instrumental variable procedure. When a system is just identified the two proce-
dures yield the same estimators (see Shapiro and Watson, 1989). If the restrictions
are placed on both the Bo and the S matrices or both Boand N( I), the ILS and 2SLS
procedures are inapplicable, in which case one could employ a method of mo-
ments approach as suggested (Bemanke, 1986). For systems that do not have a
special recursive structure or zeros appropriately placed and for overidentified
systems, Giannini (1992) develops a full-information maximum-likelihood (FIML)
estimator of the structural parameters and a formal test for overidentifying restric-
tions. Canova (1991) obtains estimates of the free parameters of the model using
a latent variable procedure.
Once behavioral economic disturbances are recovered from reduced-form VAR
innovations, summary statistics for interesting economic exercises can be col-
lected. In the semistructural VAR tradition, the typical summary statistics reported
are based on innovation accounting exercises. For example, one may be interested
in the matrix of structural MA coefficients N( f) since it gives information about
the effects of behavioral shocks on the variables of the system in the short and long
runs. Each Nj , j > I in fact describes the response of the vector X, to innovations
in a behavioral disturbance that occurred j-periods ago. The k-row of each N j
measures the responses of Xkt to innovations in the system that occurred j periods
ago, j = 0, I, .... Finally, the hth element of the kth row of N(l) measures the
cumulative effect on Xkt of the behavioral innovation with index h that occurred
j-periods ago, where j ~ 00. Operationally, the coefficients of the structural MA
representation can be obtained by setting recursively each behavioral shock of the
system equal to I at time t - j and zero afterwards and simulating the responses
of the components of X,_j+r> r = I, ... , R.
78 MACROECONOMETIUCS
To see exactly what the above expression entails, consider the case where m =2.
Then, the variance of XI' has two components: one due to the impact of its own
innovations from time t - j to time t, j = 1, 2, ... and one due to innovations in
Xu from time t - j to time t. If x 2, is shocked at time t - j and left unperturbed
afterward, we can examine how much of the variability of XI' at time t is due that
structural innovation, for all j. As (3.11) makes it clear, the variance decomposi-
tion does not add new information to the impulse response analysis but, instead,
presents it in an alternative form.
The variance decomposition and the impulse response function are, essentially,
in-sample forecasting exercises. They describe the effect on the system of a
behavioral disturbance of typical size, where by typical one usually means either
a shock or a sequence of behavioral disturbances that, on average, have occurred
during the sample period. One alternative way of summarizing the results of the
analysis is to perform a historical decomposition of x,-that is, examine the con-
tribution of each of the behavioral disturbances to the level of X,. For example, if
the dimension of X, is m = 2 and X, is stationary, then Xl,= x~,+ xl, + xi, is the
historical decomposition of x, where the first element is due to the deterministic
components, xl, =Ntl(€)u lt and xi,= Nti€)U2' where Nt( €) is the i,j element of the
matrix N'(€) = N(€)BC;IS. This type of analysis may be very useful when a re-
searcher is interested in attributing an historical episode to one particular behavioral
or policy disturbance. For example, if the VAR includes money and output and a
stock price index, it may be of interest to know if the monetary contraction of
1929, which could be represented by a series of large negative disturbances to a
monetary policy equation, is responsible for the great depression, as Friedman and
Schwartz (1963) claim, or if instead, it is due to an outlier disturbance in, say, an
equation representing the behavior of financial markets.
Initial users of VAR models reported point estimates of the impulse responses
and of the variance decomposition. As in a simple regression framework, the pre-
sentation of point estimates of the MA coefficients is not very useful unless stand-
ard errors are also provided. This is particularly important when the scope of the
economic analysis is to investigate the long-run response of endogenous variables
to behavioral shocks. Runkle (1987) shows that the Nj matrices for moderately
large j tend to be very imprecisely estimated, with the measure of dispersion
around the point estimate increasing with the lag length p of the VAR and the size
m of the system. Blanchard (1987) and Watson (1987) consider possible ways to
reduce the uncertainty around the point estimates of these statistics. They both find
VARMODELS 79
One important issue, often neglected in all the semistructural VAR literature, is the
correct number of behavioral disturbances buffeting the system. In many cases
analysts investigate the properties of bivariate systems and identification is
limited, say, to one source of supply and one source of demand disturbances.
Alternatively, one assumes that existing variables have been marginalized with
respect other potential sources of structural disturbances so that there is a one-to-
one matching of endogenous variables and shocks generating them. Both of these
approaches are obvious simplifications, and it should be clear that a misspecification
80 MACROECONOMETRICS
of the number of shocks driving the structural system is likely to create problems.
For example, if there are two sources of supply disturbances and one demand
disturbance and an investigator uses a bivariate model where only one source of
supply disturbance is identified, he is likely to confuse the impact of different
structural disturbances and bound to produce uninterpretable impulse responses.
Blanchard and Quah (1989) provide an example where these issues are thoroughly
discussed. In essence, correct identification of at least one "type" of shock is pos-
sible if and only if the distributed lag responses of the variables of the system are
sufficiently similar across types of behavioral disturbances-that is, all supply
disturbances and all demand disturbances.
While the above problem is typical of small-scale VAR models, users of
semistructural VAR models may also face the opposite problem. For example, a
semistructural VAR researcher may have in mind the implications of a RBC
model when imposing identification restrictions. However, existing RBC models
are typically driven by no more than one or two shocks, so that the vector of en-
dogenous variables may have a theoretical covariance matrix that is singular. This
observation is also relevant if a semistructural VAR researcher has in mind the
implications of a class of linear quadratic rational expectation models when im-
posing identifying restrictions. It is clear that in these situations, without further
assumptions, the behavioral disturbances are not identified. And note that the
typical assumptions used in this case, such as measurement errors or omitted
variables, may not be sufficient to recover the true sources of behavioral distur-
bances unless the variables measured with error or omitted are serially and con-
temporaneously uncorrelated with the true disturbances.
A further but related problem emerges if, instead of structural models with
shocks that have a generic behavioral interpretation, one is interested in relating
VAR disturbances to general equilibrium shocks. Because of the general equilib-
rium assumption, shocks to agents information set have both a supply-type of
effect and a demand-type effect. This is the case for example of technology or
government expenditure shocks that directly affect agents utility, which may both
increase output and demand because of the wealth effect of the shock. Structural
disturbances that represent generic aggregate demand and supply shocks may
therefore have very little relationship with these shocks and the analyses con-
ducted on the basis of these identification schemes have dubious general equilib-
rium interpretations.
There are situations however, when a VAR econometrician, observing current and
past realizations of the endogenous variables, is unable to recover the behavioral
shocks and their effects on the system. This occurs, in general, when agents in
computing their decision rule have an information set that is larger than the one
available to the econometrician. Typically, the fact that econometricians have
smaller information sets than economic agents is the rule rather than the exception,
and in many instances this is not a problem as endogenous variables are fully
revealing. That is, studying the behavior of available endogenous variables allows
a VAR researcher to discover the correct relationship between "the news" and the
endogenous variables even though the researcher never observes the news di-
rectly. Technically speaking, the endogenous variables are fully revealing of their
relationships with unobserved news when the structural moving average represen-
tation linking the endogenous variables to the behavioral shocks is fundamental-
that is, the space spanned by current and lagged x's is the same as the space
spanned by the behavioral innovations (see Quah, 1992). Among the class of
equivalent Wold MA representations, a semistructural VAR econometrician al-
ways recovers the MA that is "fundamental" (see Hansen and Sargent, 1980). This
fact has been used by West (1988) and Campbell and Deaton (1989) in rejecting
explanations of excess smoothness of consumption based on agents having larger
information sets than econometricians.
Leeper (1989), Quah (1990), Hansen, Roberds, and Sargent (1991), Hansen
and Sargent (1991), and Lippi and Reichlin (1993), however, have constructed
theoretical models where the solution for the endogenous variables are not fully
revealing of their relationship with the unobserved news. In their models the
theoretical representation of the decision rule of agents has a nonfundamental
moving average representation-that is, the space spanned by current and lagged
x's is smaller than the space spanned by behavioral innovations. In this case, a
semistructural VAR econometrician following the procedures described in this
chapter fails to recover the objects of interest of his analysis.
There does not appear to be one unifying explanation for how this phenomenon
may appear. In Leeper, nonfundamentalness is induced by the policy-reaction
function of the fiscal authority. Because tax decisions are made through a legis-
lative process that evolves slowly, economic agents may know next-period fiscal
variables when making current consumption and saving decisions. An econ-
ometrician who lacks data on unobservable tax shocks and naively identifies tax
news with the innovations in the tax process will be unable to uncover the correct
relationship between taxes and other endogenous variables because his VAR will
not be fully revealing of the relationship between news and endogenous variables.
Quah generates nonfundamentalness by assuming that agents in the economy
react to two types of shocks to labor income (permanent and transitory) while the
econometrician simply observes the realization of the process for labor income.
82 MACROECONOMETIUCS
When using a bivariate VAR process on consumption and labor income, the VAR
econometrician will not be able to recover either the permanent and transitory
shocks or their effect on consumption and labor income using current and past
values of these variables. Hansen and Sargent describe a simple linear quadratic
rational-expectations model of demand and supply, which under certain choices
of the parameters leads to a theoretical decision rule for prices and quantities that
displays a nonfundamental MA representation. They also explicitly show how the
innovations and the impulse responses recovered by the econometrician are
related to the true demand and supply shocks and their true effects on prices and
quantities. Hansen, Roberds, and Sargent show the theoretical conditions under
which a bivariate VAR model on deficit and debt can give information about
present value budget constraint relationships. Finally, Lippi and Reichlin add
technological diffusion to Blanchard and Quah's (1989) model. This addition
implies that the theoretical representation for the bivariate process for unemploy-
ment and output displays nonfundamentalness.
These examples all show that the class of theoretical models that may generate
decision rules for the endogenous variables that have a nonfundamental represen-
tation is not thin in any sense. Blanchard and Quah (1993) argue that any theoret-
ical model with fundamental representation can be transformed into a model with
nonfundamental representation without changing the underlying economic struc-
ture, by simply making use of particular orthogonal transformations (the Blaschke
factors) of their MA representation. They note that nonfundamental representa-
tions can be obtained by appropriately renormalizing the system and these re-
normalizations are observationally equivalent to fundamental ones from the point
of view of the autocovariance function of x,. An example can be informative here.
Suppose d, = 0, m = 2, 'v't and the theoretical MA has the form
with £(£., £;,) =I*CJ7. The roots of the MA polynomial are -0.25 and -0.2, which
are both less than I in modulus. In this case it is not possible to express E,'S as a
convergent linear combination of current and past x,' s. Therefore the above system
is not in a fundamental MA form because the space spanned by the E'S is larger
than the space spanned by the x's. It is easy to check that the x's generated from
(3.12) and the x's generated by
can be expressed in terms of current and past x,'s with coefficients which are
square summable-that is, if l1ajX,_j = u, the coefficients aj satisfy ~ja] < 00. The
system (3.13) is in a fundamental (Wold) MA form. Note that to go from (3.12)
to (3.13) we have postmultiplied the MA coefficients by matrices like iJ =
(e-a;)]
[ aie
diag 1 _ where ajis the ith root of the original model and ('Xi its complex
conjugate, which have the property that Btl =I. Despite this generic observation-
ally equivalent result, Lippi and Reichlin (1992) observe that there are less trivial
nonfundamental transformations than the one obtained by Blanchard and Quah
(1993). These transformations, which still have the same autocovariance function
for x" have economic meaning and can be obtained by slightly changing the
economic structure of the theoretical model.
The implications of these findings are, at the present, not fully explored. It is
clear, however, that when the theoretical model delivers nonfundamental repre-
sentations, VAR analysis fails to capture the news to economic agents and the
their impact on the endogenous variables. One implication of these results is that
one should at least check whether the results are sensitive to the assumption of
fundamental representation implicit in the VAR analysis.
with memory that can go arbitrarily far in the past. This feature therefore implies
that discrete-time "news" are weighted averages of the news hitting agents' infor-
mation set over a possible large interval of time and, as such, need not to have
much relationship with the true innovations at each t. Second, evidence of Granger
causality in discrete-time data need not bear a relationship with the Granger cau-
sality existing in continuous time, unless the VAR econometrician finds no evi-
dence of causality. Third, and as a consequence of the above, the impulse response
analysis conducted in discrete time may have little relationship with the true
continuous time impulse responses.
An Example
To demonstrate in practice some of the problems that may emerge with identifi-
cation and interpretation of the reSUlts, I next present an example where it is
possible to quantify the significance of certain assumptions and to measure the
distortions implied by particular identification schemes.
I consider a version of the bivariate model studied by Quah (1990) where
disposable income is a random process that has two components and agents take
it as given. I assume that agents hold rational expectations and that they are
interested in choosing the best possible path for consumption over time. Depend-
ing on the assumptions I will make, the two components of income have different
interpretations.
The first one, suggested by Quah (1990), is that one component represents the
permanent part of income and the second one the transitory part. The two com-
ponents are theoretically identified by the assumption that the disturbances are
uncorrelated and that, by definition, the transitory part is a process that has no
long-run effects on the level of income. Note that in this interpretation, the two
components are simply indices that need not have a unique economic interpreta-
tion. They can be identified with supply and demand shocks (as, for example, in
Blanchard and Quah, 1989) or with real and monetary shocks (as in Robertson and
Wickens, 1992).
The second interpretation identifies the first component with the result of pri-
vate agents' behavior and the second component with policy behavior. For exam-
ple, the second component of income may be driven by a shock that can be called
the stochastic income tax rate while the first component can be driven by technol-
ogy shocks. A model with this type of interpretation is provided by Leeper (1989).
These two components may be contemporaneously uncorrelated if government
policy is exogenous to the economy, contemporaneously correlated if the policy
authority reacts to the current state of the economy or correlated with a lag if,
VARMODELS 85
within the unit of time of the model (such as the quarter), the policy authority does
not see the current state of the economy.
Within these two interpretations we will consider two cases-one where the
theoretical model generates a moving average representation that is fundamental
for the joint consumption-income process and one where the representation is not
fundamental.
Formally speaking, the process for income will be of the form
Yt =Ylt + Y2t>
Ylt = a,(f)u lt ,
Y2t =a2(f)u 2t • (3.14)
ai f) are polynomials in the lag operator and E(up;) = I where u; = [u lt , u2t J'. In
Model I, I assume that a 1(0) # 0, that a 2(0) = 0 where aj(O) :; ajO' and that I =
diag(O'I' 0'2) so that ~t does not instantaneously Granger cause u lt • In Model II, I
will assume that aiel) #0 and that all) = 0 where ap):; Ijaj and that I= diag(O'"
0'2) so that YII has a unit root but Y2t does not.
Following Quah (1990), if agents have rational expectations and maximize
utility using a fixed interest rate to compute future annuities, the resulting theoreti-
cal bivariate moving-average representation for income and consumption will be
of the form
(3.15)
where f3 is the agents' discount factor and a/{3) :; aj ( f = f3). Note that if f3 < I,
Models I and IT generate a nonfundamental MA representation, while if f3> I, both
models generate fundamental representations (so that in total we have four differ-
ent data-generating processes (DGP) ). For all the experiments with Model I al( f)
= 0.6 - 0.3f + 0.2f 2 and a2(f) = 0.0 + 0.2f - 0.5f 2• For all the experiments with
Model IT al(f) = 0.6 - 0.3f + 0.2f 2 and a2(f) = 1.0 + 0.2f - 0.5f2• The discount
factor beta can take two values f3= 0.96 or f3= 1.04 where the last value is selected
following the considerations contained in Kocherlakota (1990).
We simulate data out (3.15) assuming that the time interval of the model is a
quarter. For each of the four bivariate processes generated, we estimate a VAR(4),
apply two identification schemes, and attempt to interpret the "structural innova-
tions" and the corresponding structural impulse responses. The first identification
scheme is the same as Blanchard and Quah (1989), where one of the shocks has
no long-run effect on the variables of the system. The second is one where one of
the shocks has no contemporaneous impact on the variables of the system. Figures
86 MACROECONOMETRICS
3.1 and 3.2 present the results. The first and the third rows of both figures report
the responses of ay, while the second and the fourth rows the responses of ac for
each of the two processes. In Figure 3.1 f3 = 1.04, in Figure 3.2 f3 =0.96. The first
column presents the theoretical responses; the second the responses generated
with an identification scheme that places zeros on the contemporaneous correla-
tion among shocks; the third the responses generated with a Blanchard and Quah
decomposition. The solid line in each box is the response to shock 1, and the
dashed line is the response to shock 2.
The results of the simulation exercise display some interesting features. First,
when the DGP is a fundamental MA process, even when the choice of identifica-
tion is correct (as it is the case of identification I when DGPI is the true process
and identification 2 when DGP2 is the true process), there may be some error in
determining the size of the initial impact of the shocks (there is a downward bias
of about 50 to 80 percent), in detecting the overall size of the total multipliers, and,
in the case of DGP2 even, in recovering the overall shape of the responses. For
example, when we use DGP2 and identification 2, the initial impact of U. , on ay
and ac (the solid line) has an incorrect sign and the shape of the responses do not
correspond to the true ones. There are several reasons for why these results may
emerge. First, we have used a VAR(4) to approximate a VAR(oo) so the model is
misspecified. Second, the sample size may be too small to recover the correct
responses. Third, since the size of the responses is in many cases close to zero, it
may be, that once standard errors are taken into account, some sign reversal may
not be significant. When the identification scheme is incorrect, several problems
can be noted. The most evident problem is that the effect of the two shocks is
reversed for both variables so that if one uses the wrong identification scheme he
is likely to confuse the correct source of disturbances.
When the DGP is a nonfundamental MA, regardless of the exact DGP and
identification scheme used, substantial errors emerge. First, in many cases shocks
hitting consumption are incorrectly identified and their effect reversed. Second,
especially when DGP2 is the true process and identification 2 is used, the sign of
the initial response is incorrect. Third, even when the sign is correct, the magni-
tude of the impact is often 80 percent smaller than the true one. Finally, the shape
of the responses to the two shocks is often incorrect.
Two conclusions can be drawn from the above exercise. First, when the DGP
is a fundamental MA process, the use of an incorrect identification scheme may
lead a researcher to confuse sources of disturbance, to underestimate the impact
of the shocks, and, in some cases, to have an incorrect picture of the propagation
of shocks over time. Second, when the DGP is a nonfundamental MA, both
identification schemes provide a very distorted picture of the propagation mech-
anism of shocks in a bivariate system. This may lead a researcher to draw incorrect
conclusions regarding the sources of shocks and their impact over time.
87
Identification I , 2
ldenn·cot·on
"
0....
I
"
"
-.... " lI\f\....
"
\
0.00
" ' ·V
"
!
_... J...,-""",,~~~~~~~,....,.....J
7 12 17 -....
Income
O.lXXJIi..-----------,
0.0027
, I
I
I
0.IX)l8 I
I
,,
2 7 12
CClf'lSU'T1)tion
0:0; 0.1'
021 ~
, .1\
,~ .
0.00
0.12
0.00
'VV ~
·.1'
V
-.12
-21
V
-.32
88
T"",' Nonf""*>nentol Mil
0.75
0.50
-.25
, I
"~
-.50L.-"';,~~~~~~~-1
17 17
"02ll
0.010
0.000
,
,,
-.010 r
-.02ll
.00 Tra."'T~,~~,~",",,"'"\Tf'A,
2 7 12 17
Figure 3.2
VARMODELS 89
Conclusions
The purpose of this chapter was to describe the main features of the unrestricted
VAR approach to macroeconometrics and of semistructural exercises conducted
with VARs and to discuss their meaning and interpretation. The emphasis was
primarily on the economic aspects and on the interpretation of the results of the
procedure rather than on statistical issues surrounding the approach (which are
more thoroughly discussed in Lutkepoh1 (1991); Hamilton, 1994; Canova, 1993;
and Watson, 1994). We have attempted to compare and contrast different ap-
proaches to specification and testing and have discussed differences with other
existing methodologies for applied research.
Criticisms of the unrestricted VAR methodology discussed here appeared at
two levels: at the stage of describing data interdependences and at the stage of un-
dertaking useful policy analyses. Some authors (see Runkle, 1987; Ohanian, 1988;
Spencer, 1989) have raised doubts about the usefulness of the methodology to
robustly characterize the dynamics of the data (a result contradicted by the recent
findings of Todd, 1990). In addition Maravall (1993) stressed that VAR models
fit to seasonally adjusted data are unable to capture the dynamics of the data be-
cause seasonal adjustment procedures induce noninvertible MA components in
the data.
The major bulk of the criticisms has focused, however, on the meaning and on
the limitations of hypothesis testing and policy analysis with VARs. Attacks were
mounted against its "atheoretical" underpinning (see Cooley and Leroy, 1985;
VARMODELS 91
Leamer, 1985; Blanchard and Watson, 1986), on the type of policy interventions
that were reasonable in light of the Lucas critique, on the way they were en-
gineered (see Cooley and Leroy, 1985), and, in general, on the usefulness of pro-
viding policy advice with VARs (see Sargent, 1984). We argued that semistructural
VAR analyses do have theoretical links, even though they are generic and refer to
a large class of models. We have also argued that they can be used to provide
answers to some questions that are of interest for policy makers and economists,
but we also discuss their major limitations. We show that there are circumstances
where the semistructural VAR econometrician cannot recover relevant news to
agents' infonnation sets and describe their impact on the variables of the system.
These problems, however, are not unique to the methodology described in this
chapter. Instead, they are typical of all the procedures that do not heavily rely on
economic theory to put structure on the data and instead "let the data speak."
Unfortunately, and in addition to the above problems, in many situations the data
does not speak loud enough for the evidence presented to be conclusive. This,
however, should not be regarded as a failure of the methodology but instead must
be regarded as a weakness of the data. More precise models of economic behavior
can overcome most of the deficiencies faced by the semistructural VAR analysis.
However, very few economists really believe that our highly stylized economic
models can be taken to be the true data-generating process for the actual data, and
it is symptomatic that detailed models of economic behavior are almost always
rejected by fonnal statistical testing. When nonrejections occur, it is often because
of small samples or inefficient estimation techniques.
Whether the unrestricted semistructural VAR approach is more or less fruitful
than approaches that rely on strongly fonnulated, and possibly false, theoretical
models is still an open issue that deserves further study. We conjecture that the
answer will not be univocal and will depend on the questions the researchers are
interested in asking.
Acknowledgments
References
Akaike, H. (1974). "A New Look at the Statistics Model Identification." IEEE Transactions
on Automatic Control AC-19, 716-723.
Baillie, R.T. (1987). "Inference in Dynamic Models Containing Surprise Variables." Journal
of Econometrics 35, 101-117.
Bernanke, B. (1986). "Alternative Explorations of the Money Income Correlation." Carnegie
Rochester Conference in Public Policy 25, 49-100.
Beveridge, S., and C. Nelson. (1981). "A New Approach to Decomposition of Economic
Time Series into Permanent and Transitory Components with Particular Attention to
Measurement of Business Cycle." Journal of Monetary Economics 7, 151-174.
Blanchard, O. (1987). "Comment to Runkle's Vector Autoregression and Reality." Journal
of Bussiness and Economic Statistics 5, 449-451.
Blanchard, O. (1989). "A Traditional Interpretation of Macroeconomic Fluctuations."
American Economic Review 79, 1146-1164.
Blanchard, 0., and D. Quah. (1989). 'The Dynamic Effects of Aggregate Demand and
Supply Disturbances." American Economic Review 79,655-673.
Blanchard, 0., and D. Quah. (1993). "Fundamentalness and the Interpretation of Time
Series Evidence: A replay to Lippi and Reichlin." American Economic Review 83(3),
653-658.
Blanchard, 0., and M. Watson. (1986). "Are Business Cycles all Alike?" In Gordon, R.
(ed.), The American Business Cycle. Chicago: University of Chicago Press.
Brock, W., and C. Sayers. (1988). "Is the Business Cycle Characterized by Deterministic
Chaos?" Journal of Monetary Economics 22, 71-90.
Campbell, J., and A. Deaton. (1989). "Why Is Consumption so Smooth?" Review of Eco-
nomic Studies 56, 357-373.
Canova, F. (1991). "The Sources of Financial Crises: Pre and Post Fed Evidence." Inter-
national Economic Review 32, 689-713.
Canova, F. (1993). "Modelling and Forecasting Exchange Rates with Bayesian Time Vary-
ing Coefficient Model." Journal of Economics Dynamic and Control 17,233-261.
Canova, F. (1994). "VAR Models: Specification, Estimation, Inference and Forecasting."
In H. Pesaran and M. Wickens (eds.), Handbook of Applied Econometrics. Oxford:
Basil Blackwell.
Canova, E, M. Finn, and A. Pagan. (1993). "Evaluating a Real Business Cycle Model." In
C. Hargreaves (ed.), Nonstationary Time Series Analysis and Cointegration. Oxford:
Oxford University Press.
Clements, M.P., and G.E. Mizon. (1991). "Empirical Analysis of Macroeconomic Time
Series: VAR and Structural Models." European Economic Review 35,887-917.
Cooley, T., and S. Leroy. (1985). "Atheoretical Macroeconometrics: A Critique." Journal
of Monetary Economics 16,283-308.
Dickey, D., and W. Fuller. (1979). "Distribution of the Estimators for a Autoregressive Time
Series with a Unit Root." Journal ofthe American Statistical Association 74, 427-431.
Danthine, J.P. (1992). "Calibrated Macro Models: What and What For?" Manuscript,
University of Lousanne.
VARMODELS 93
Dejong, D., and Whiteman, C. (1991). "Reconsidering Trends and Random Walks in
Macroeconomic Time Series." Journal of Monetary Economics 28(2),221-254.
Doan, T. (1994). RATS Manual, Version 4.10. Evanston, II.
Doan, T., R. Linerman, and C. Sims. (1984). "Forecasting and Conditional Projections
Using Realist Prior Distributions," Econometric Reviews 3, 1-100.
Engle, R. (1982). "Autoregressive Conditional Heteroskedasticity with Estimates of the
Variance of the UK Inflation." Econometrica 50, 987-1006.
Engle, R., and C. Granger. (1987). "Co-Integration and Error Correction: Representation,
Estimation and Testing." Econometrica 55,251-276.
Engle, R., D. Lilien, and R. Robbins. (1987). "Estimating Time Varying Risk Premia in the
Term Structure: The ARCH-M ModeL" Econometrica 55, 391-408.
Evans, G. (1989). "Output and Unemployment Dynamics in the United States: 1950-
1985," Journal of Applied Econometrics 4,213-237.
Fair, R. (1988). "VAR Models as Structural Approximations," Discussion Paper 856, Cowles
Foundation.
Friedman, M., and A. Schwartz. (1963). A Monetary History of the United States,1867-
1960. Princeton, NJ: Princeton University Press.
Gali, J. (1992). "How Well Does the ISLM Model Fit Postwar U.S. Data?" Quarterly
Journal of Economics 107, 709-735.
Gali, J., and M. Harnmour. (1991). "Long Run Effects of Business Cycles," Manuscript,
Columbia University.
Gallant, R., D. Hsieh, and G. Tauchen. (1991). "On Fining Recalcitrant Series: The Pound!
Dollar Exchange Rate, 1974-1983," In W. Barnen, J. Powell and G. Tauchen, (eds.),
Nonparametric and Semiparametric Methods in Econometrics and Statistics. Proceed-
ings of the fifth International Symposium in Economic Theory and Econometrics.
Cambridge: Cambridge University Press.
Geweke, J. (1982). "The Measurement of Linear Dependence and Feedback Between
Multiple Time Series," Journal of the American Statistical Association 77, 304-313.
Giannini, C. (1992). Topics in Structural VARs. Berlin: Springer and Verlag.
Granger, c., and Anderson. (1978). Introduction to Bilinear Time Series. Goningen, Ger-
many: Vandenhoeck and Ruprecht.
Griffiths, W., and H. LutkepoW. (1990). "Confidence Interval for Impulse Responses from
VAR Models: A Comparison of Asymptotic Theory and Simulation Approaches,"
Manuscript, University of New England.
Hamilton, 1. (1989). "A New Approach to Economic Analysis of Nonstationary Time
Series and the Business Cycles," Econometrica 57, 357-384.
Hamilton, 1. (1994). Applied Time Series Analysis. Princeton, NJ: Princeton University
Press.
Hansen, L. (1982). "Large Sample Properties of Generalized Methods of Moment Estima-
tors." Econometrica 50, 1029-1054.
Hansen, L., W. Roberds, and T. Sargent. (1991). "Time Series Implications of Present
Value Budget Balance and of Martingale Models of Consumption and Taxes," In L.
Hansen and T. Sargent (eds.), Rational Expectations Econometrics. Boulder, Co:
Westview Press.
94 MACROECONOMETRICS
Hansen, L., and T. Sargent. (1980). "Fonnulating and Estimating Linear Rational Expec-
tations Models." Journal of Economic Dynamics and Control 2, 1-46.
Hansen, L., and T. Sargent. (1991). "Two Difficulties in Interpreting Vector Autoregres-
sions." In L. Hansen and T. Sargent (eds.), Rational Expectations Econometrics. Boulder,
Co: Westview Press.
Hansen, L., and K. Singleton. (1983). "Stochastic Consumption, Risk Aversion and the
Temporal Behavior of Asset Returns." Journal of Political Economy 91, 249-265.
Harvey, A.C. (1985). "Trends and Cycles in Macroeconomics." Journal of Business and
Economic Statistics 3, 216-227.
Hendry, D., and G. Mizon. (1990). "Evaluating Dynamic Econometric Models by Encom-
passing the VAR." In P.C.B. Phillips and V.B. Hall (eds.), Models, Methods and Ap-
plications of Econometrics: Essays in Honor of Rex Bergstrom. Oxford: Oxford
University Press.
Hsiao, C. (1981). "Autoregressive Modelling and Money-Income Causality Detection."
Journal of Monetary Economics 7, 85-106.
Johansen, S. (1988). "Statistical Analysis of Cointegrating Vectors." Journal of Economic
Dynamics and Control 12, 231-254.
Johansen, S., and K. Juselius. (1990). "Maximum Likelihood Estimation and Inference on
Cointegration-with Application to the Demand for Money." Oxford Bulletin of Eco-
nomics and Statistics 52, 169-210.
Judge, G., W. Griffith, R. Hill, H. Lutkepohl, and T. Lee. (1980). The Theory and Practice
of Econometrics (2nd ed.). New York: Wiley.
Keating, 1. (1990). "Identifying VAR Models under Rational Expectations." Journal of
Monetary Economics 25, 453-476.
Kirchgassner, G. (1991). "Comments on M.P. Clements and G.E. Mizan: Empirical Analy-
sis of Macroeconomic Time Series: VAR and Structural Models." European Economic
Review 35, 918-922.
King, R., C. Plosser, and S. Rebelo. (1988). "Production, Growth and Business Cycles, I."
Journal of Monetary Economics 21, 309-342.
King, R., C. Plosser, J. Stock, and M. Watson. (1991). "Stochastic Trends and Economic
Fluctuations." American Economic Review 81, 819-840.
Kocherlakota, N. (1990). "On the Discount Factor in Growing Economies." Journal of
Monetary Economics 25, 43-48.
Koopmans, T. (1963). "Identification Problems in Economic Model Construction." In W.
Hood and TJ. Koopmans (eds.), Studies in Econometric Methods. New Haven, Cf:
Yale University Press.
Kydland, F., and E. Prescott. (1982). "Time to Build and Aggregate Fluctuations."
Econometrica 50, 1345-1370.
Kydland, F., and E. Prescott. (1991). ''The Econometrics of the General Equilibrium Ap-
proach to Business Cycles." Scandinavian Journal of Economics 93, 161-178.
Leamer, E. (1978). Specification Searches and ah-Hoc Inference with Nonexperimental
Data. New York: Wiley.
Leamer, E. (1985). "Vector Autoregression for Causal Inference?" Carnegie Rochester
Conference Series on Public Policy 22, 255-304.
VARMODELS 95
Leeper, E. (1989). "Policy Rules, Infonnation and Fiscal Effects in a 'Ricardian' Model."
International Finance Discussion Paper 360, Board of Governors of the Federal Reserve
System.
Leeper E., and R. Gordon. (1992). "In Search for the Liquidity Effects." Journal of
Monetary Economics 29, 341-371.
Lippi, M., and L. Reichlin. (1992). "VAR Analysis, Non-Fundamental Representation,
Blaschke Matrices." Manuscript, OCFE.
Lippi, M., and L. Reichlin. (1993). "A Note on Measuring the Dynamic Effects of Aggre-
gate Demand and Supply Disturbances." American Economic Review 83(3), 644-652.
Littennan, R. (1980). "Techniques for Forecasting Vector Autoregressions." Ph.D. thesis,
University of Minnesota.
Littennan, R. (1984). "Forecasting and Policy Analysis with Bayesian Vector Autoregression
Models." Quarterly Review (Federal Reserve Bank of Minneapolis) 8 (Fall), 30-41.
Littennan, R. (1986). "Specifying Vector Autoregressions for Macroeconomic Forecast-
ing." In P. Goel and A. Zellner (eds.), Bayesian Inference and Decision Techniques.
Amsterdam: North Holland.
Littennan, R., and L. Weiss. (1985). "Money, Real Interest Rates and Output: A Reinter-
pretation of Postwar U.S. Data." Econometrica 52, 129-156.
Liu, T. (1960). "Underidentification, Structural Estimation and Forecasting." Econometrica
28, 855-865.
Long, J., and C. Plosser. (1983). "Real Business Cycles." Journal of Political Economy 91,
1345-1370.
Luppoletti, W., and R. Webb. (1986). "Defining and Improving the Accuracy of Macro-
economic Forecasts: Contributions from a VAR Model." Journal of Business 59, 263-
285.
LutkepoW, H. (1986). "Comparison of Criteria for Estimating the Order of a Vector Auto-
regressive Process." Journal ofTime Series Analysis 6,35-52; and "Correction," 8, 373.
LutkepoW, H. (1991). Introduction to Multiple Time Series Analysis. New York: Springer
and Verlag.
Maravall, A. (1993). "Stochastic Linear Trends: Models and Estimators." Journal of
Econometrics 56,5-37.
McNees, S. (1986). "Forecasting Accuracy of Alternative Techniques: A Comparison of
U.S. Macroeconomic Forecasts" (with comments). Journal of Business and Economic
Statistics 4, 23-28.
McNees, S. (1990). "The Role of Judgement in Macroeconomic Forecasting Accuracy."
International Journal of Forecasting 6, 287-299.
Nelson, and C. Plosser. (1982). "Trends and Random Walks in Macroeconomic Time
Series." Journal of Monetary Economics 10, 139-162.
Nickelsburg, G. (1985). "Small Sample Properties of Dimensionality Statistics for Fitting
VAR Models to Aggregate Economic Data." Journal of Econometrics 28, 183-192.
Nicholls, D.F., and B.G. Quinn. (1982). Random Coefficient Autoregressive Models: An
Introduction. New York: Springer and Verlag.
Ohanian, L. (1988). "The Spurious Effect of Unit Roots on Vector Autoregressions: A
Monte Carlo Study." Journal of Econometrics 39, 251-266.
96 MACROECONOMETIUCS
Perron, P. (1989). "The Great Crash, the Oil Price Shock and the Unit Root Hypothesis."
Econometrica 57, 1361-1401.
Phillips, P.C. (1988). "Optimal Inference in Cointegrated Systems." Econometrica 55,
277-301.
Phillips, P.C., and P. Perron. (1986). "Testing for Unit Roots in Time Series Regressions."
Biometrika 70, 335-346.
Quah, D. (1990). "Pennanent and Transitory Movements in Labor Income: An Explanation
Based on Excess Smoothness in Consumption." Journal of Political Economy 98, 449-
475.
Quah, D. (1991). "Identifying Vector Autoregressions: A Discussion of P. Englund, A.
Vredin and A. Warne: Macroeconomic Shocks in Sweden 1925-1986." Manuscript,
London School of Economics.
Quah, D. (1992). "Lecture Notes in Macroeconometrics." Manuscript, London School of
Economics.
Quah, D., and T. Sargent. (1993). "A Dynamic Index Model for a Large Cross Section."
In J. Stock and M. Watson (eds.), New Research on Business Cycles, Indicators and
Forecasting. Chicago: University of Chicago Press.
Reinsel, G. (1983). "Some Results on Multivariate Autoregressive Index Models."
Biometrika 70, 145-156.
Robertson, D., and M. Wickens. (1992). "Measuring Real and Nominal Macroeconomic
Shocks and Their International Transmission Under Different Monetary Systems."
Working paper, London Business School.
Rozanov, Y. (1967). Stationary Random Processes. San Francisco: Holden Day.
Runkle, D. (1987). "Vector Autoregression and Reality." Journal of Business and Eco-
nomic Statistics 5, 437-454.
Sargent, T. (1979). "Estimating Vector Autoregressions Using Methods Not Based on
Explicit Economic Theory." Quarterly Review (Federal Reserve Bank of Minneapolis),
3 (Summer), 8-15.
Sargent, T. (1984). "Autoregression, Expectations and Advice." American Economic Re-
view, Papers and Proceedings 74, 408-415.
Sargent, T. (1986). Macroeconomic Theory (2nd ed.). New York: Academic Press.
Sargent, T., and R. Lucas. (1980). Rational Expectations and Econometric Practice.
Minneapolis, MN: University of Minnesota Press.
Sargent, T., and C. Sims. (1977). "Business Cycle Modelling Without Pretending to Have
Too Much A Priori Economic Theory." In C. Sims (ed.), New Methods in Business
Cycle Research. Minneapolis, MN: Federal Reserve Bank of Minneapolis.
Schwarz, G. (1978). "Estimating the Dimension of a Model." Annals of Statistics 6, 461-
464.
Shapiro, M., and M. Watson. (1989). "Sources of Business Cycle Fluctuations." NBER
Macroeconomic Annual (pp. 111-148). Cambridge, MA: MIT Press.
Sims, C. (1972). "Money, Income and Causality." American Economic Review 62, 540-
553.
Sims, C. (1980a). "Macroeconomics and Reality." Econometrica 48, 1-48.
Sims, C. (1980b). "A Comparison of Interwar and Postwar Business Cycles: Monetarism
Reconsidered." American Economic Review Papers and Proceedings 70, 250-257.
VARMODELS 97
Sims, C. (1982). "Policy Analysis with Econometric Models." Brookings Papers of Eco-
nomic Activity 2, 107-152.
Sims, C. (1986). "Are Forecasting Models Usable for Policy Analysis." Quarterly Review
(Federal Reserve Bank of Minneapolis), 10 (Winter), 2-16.
Sims, C. (1987). "Comment to Runkle's Vector Autoregression and Reality." Journal of
Business and Economic Statistics 5, 443-449.
Sims, C. (l988a). "Identifying Policy Effects." In R. Bryant, D. Henderson, G. Holtman,
P. Hooper, and A. Symansky (eds.), Empirical Macroeconomics for Interdependent
Economies (pp. 305-321). Washington DC: Brookings Institution.
Sims, C. (I 988b). "Bayesian Prospective on Unit Roots Econometrics." Journal of Eco-
nomic Dynamic and Control 12, 436-474. .
Sims, C. (l989b). "A Nine Variable Probabilistic Model of the U.S. Economy." Discussion
Paper 14, Institute for Empirical Macroeconomics, Federal Reserve of Minneapolis.
Sims, c., J. Stock, and M. Watson. (1990). "Inference in Linear Time Series Models with
Some Unit Roots." Econometrica 58, 113-144.
Sims, C., and H. Ulhig. (1991). "Unit Rooters: An Elicopter Tour." Econometrica 59(6),
1591-1600.
Spencer, D. (1989). "Does Money Matter? The Robustness of the Evidence from Vector
Autoregressions." Journal of Money. Banking and Credit 21, 442-454.
Stock, J., and M. Watson. (1988). "Testing for Common Trends." Journal ofthe American
Statistical Association 83, 1097-1107.
Stock, J., and M. Watson. (1989). "Interpreting the Evidence on Money-Income Causality."
Journal of Econometrics 40, 161-181.
Taylor, 1. (1980). "Aggregate Dynamics and Staggered Contracts." Journal of Political
Economy 88, 1-24.
Theil, H. (1971). Principles of Econometrics. New York: Wiley.
Todd, R. (1984). "Improving Economic Forecasting with Bayesian Vector Autoregression."
Quarterly Review (Federal Reserve Bank of Minneapolis), 8 (Fall), 18-29.
Todd, R. (1990). "Vector Autoregression Evidence on Monetarism: Another Look at the
Robustness Debate." Quarterly Review (Federal Reserve Bank of Minneapolis), 14
(Spring), 17-37.
Watson, M. (1986). "Univariate Detrending Methods with Stochastic Trends." Journal of
Monetary Economics 18,49-75.
Watson, M. (1987). "Comment to Runkle's Vector Autoregression and Reality." Journal
of Business and Economic Statistics 5,451-453.
Watson, M. (1994). "VAR and Cointegration." In R. Engle and D. McFadden (eds.),
Handbook of Econometrics (Vol. 4).
Wickens, M. (1992). "V AR Analysis in the Presence of Cointegration." Working paper,
London Business School.
Zellner, A. (1962). "An Efficient Methods for Estimating Seemingly Unrelated Regressions
and Tests for Aggregation Bias." Journal of the American Statistical Association 57,
348-368.
Commentary on Chapter 3
Masao Ogaki
Fabio Canova has provided us with an excellent discussion of the VAR method-
ology's relationship to economic models. In this comment, I focus on the identi-
fication issue. Canova defines the question of identification in the following way:
under what conditions is knowledge about reduced-form VAR parameters and
innovations sufficient to recover objects that may be of interest to economic analy-
sis? Canova offers a general discussion on this issue.
In order to understand the relationship between conditions of identification and
economic models, it is helpful to have a concrete example of an economic model.
For this purpose, a simple linear rational expectations model that imposes restric-
tions on the VAR representation is presented here. Using this example, I discuss
some important conditions for identification under the VAR methodology and
compare the VAR methodology with other competing approaches.
P, = E[b(PI+1 + d +
l I) II,l, (1)
where b is the constant real discount rate and E(· I I,) is the mathematical expec-
tation operator conditioned on the information set II economic agents in period t.
Solving (1) forward and imposing the no bubble condition, we obtain the present-
value formula:
(2)
We now derive restrictions on the VAR representation for P, and d, implied by (1)
and (2). We consider two cases, depending on whether d, is assumed to be
covariance stationary or unit root nonstationary.
99
100 MACROECONOMETIUCS
Case 1: Stationary d,
Assume that d, is covariance stationary with mean zero, so that it has a Wold
moving average representation
d, = a(L)v" (3)
where
(7)
and E(w, I HI) = O. Because E(· I H,) is the linear projection operator onto H"
where ~(L) = ~l + ~2L. ... Following Hansen and Sargent (1980, app. A), we
obtain the restrictions imposed by (8) on ~(L) and tfJ(L):
~ bL-I(a(L) - a(b» bL-1(1- tfJ-l(b)tfJ(L» (9)
(L) = 1 _ bL-I tfJ(L) = 1 _ bL-I
Then (9) is used to show that ~(L) is a finite-order polynomial and to give a
explicit formula for the coefficients for ~(L) (see West, 1987, for the formula,
COMMENTARY ON CHAPTER 3 101
which is based on Hansen and Sargent, 1980, and on West, 1989, for detenninistic
terms when d(t) has a nonzero mean). Thus
P, = bId, + ... + bqd,_q+1 + W, , (11)
where b;' s are functions of b and ~;' s.
For simplicity, we started out with an assumption that the econometrician's
information set is H, = {d" d'_l' d'_2'" .}. In order to obtain the VAR represen-
tation for (dt+l' P,), we need to assume that the econometrician's information set
is H~= {d" d'_I' d'_2' ... , P,_I' P'-2' P,-3" .. }. For simplicity, we now assume that
that H, and H~ has the same informational contents for d'+1 and P,: E(d'+1 I H,) =
E(dr+1 I H~), E(p, I H,) = E(p, I H~), so that (10) and (11) are the VAR represen-
tation. Without this simplifying assumption, the only thing we would need to do
is to apply Hansen and Sargent's (1980) formula to H~. Then we would obtain
very similar restrictions as (9) except that ~(L) and a(L) become 2 by 2 matrices
and that both current and lagged d,'s and lagged p,'s appear in (10) and (11) as
regressors. Hence the following discussions are not affected by this simplifying
assumption. Thus (9) gives the restrictions on the VAR representation consisting
of (10) and (11).
Econometric Methods
In this section, econometric methods that impose and test the nonlinear restrictions
discussed in the previous section are described. I focus on Hansen and Sargent's
102 MACROECONOMETRICS
Case 1: Stationary d t
Let ZI, be a vector of random variables in H t • For example, ZIt = (d" ... d'_q+I)' The
unknown parameters band qJ;' s can be estimated by applying the GMM to
orthogonality conditions E(ZI,V,+I) =0 and E(zl,w,) =0 in the econometric system
consisting of (10) and (11).
Let Z2' be a random variable in I,-say, d,-and
(12)
Then (1) implies another orthogonality condition E(Z2'U ITI) = O. This orthogonality
condition can be used to estimate b. West (1987) forms a specification test a la
Hausman (1978) by comparing the estimate of b from (12) with the estimate of
b from (10) and (11). For this purpose, West forms a Wald test in the system
consisting of (10), (11), and (12) without the restrictions (9) imposed. Another
method to form West's specification test is to form a Lagrange multiplier test or
a likelihood ratio type test, which will require estimation constrained by the re-
strictions (9). This method may be preferable because of small sample problems
with the Wald test for nonlinear restrictions (see Ogaki, 1993, sec. 7, for discus-
sions about these tests).
In this case, the GMM can be applied to the econometric system consisting of
(10') and (11') to estimate band qJ;'s. A complication exists for (12). It should be
noted that u, in (12) is stationary, so that (12) implies that P, and Pt+1 + dt+ t are
cointegrated with a cointegrating vector (I, -b) in Engle and Granger's (1987)
terminology. In order to see this, note that
(13)
Because d,.,; - d, is stationary for any i, the right side of (13) is stationary (as long
as the stationary random variables in I, are enough to form optimum forecasts for
d,.,; - d,). Hence p,and d,are cointegrated with a cointegrating vector [I, - b (1-
b)]' as shown by Campbell and Shiller (1987). This implies that P, is unit root
nonstationary and that
COMMENTARY ON CHAPTER 3 103
is stationary.
There exist at least three alternative methods to deal with (12). West (1987)
notes that his instrumental variable estimator for b from (12) has an asymptotic
normal distribution when d, has nonzero drift (equation (10') above ignores nonzero
drift). This is because P, and P,+1 + d +1 satisfy the deterministic cointegration
'
restriction in Ogaki and Park's (1990) terminology and because there is only one
regressor in (12). Hence West's (l988b) results apply to this case. West forms a
Wald test that compares his estimate of b from (12) and his estimate of b from
(10') and (II'), using the joint normal distributions for the estimators from the
system consisting of (10'), (II'), and (12).
Another method, which is used in Cooley and Ogaki (1995) for a different
model, can be applied in more general cases. Because the estimator for b from (12)
converges faster than GMM estimators from (10') and (II') because of the super
consistency, a two-step procedure is possible. In the first step, b is estimated from
(12) by a cointegrating regression. It is preferable to use one of asymptotic effi-
cient estimators than OLS estimators. In the second step, the estimate of b from
(12) is used as the true value for the GMM estimation of band tPi' s from (10') and
(11'). Because the estimator of b from (12) converges faster than the GMM esti-
mators, this two-step procedure does not affect asymptotic distributions of the
GMM estimators.
Another method is to first difference (l2}-
(12')
and use the orthogonality condition E(Llu +J I I,_I) = O. This may be somewhat
'
simpler, but we lose information from stochastic trends by first differencing. The
econometrician needs to be careful because Llu'+1 is overdifferenced. For example,
a constant is not a valid instrument because the resulting long run covariance
matrix for the GMM disturbance would be singular.
Some remarks are in order.
• Hansen and Sargent's method described above does not require an assump-
tion that d, is exogenous. Relation (10) or (10') is obtained from the assump-
tion that d, is covariance stationary and that its Wald representation is
invertible.
• For the econometric system consisting of (10) and (11 ) (or (10') and (11') ),
random variables in H, can be used as instruments, but the variables in I, that
are not in H, are not valid instruments by construction.
104 MACROECONOMETIUCS
• Because U'+I in (12) is in 1'+1 and V,+( in (10) is in Ht+(, U, and V, are serially
uncorrelated (see Ogaki, 1993, sec. 6, for related discussions). However,
W, in (11) is not necessarily in H,+(. Hence W, has unknown order of serial
correlation.
Comparing Methods
Compared with a method that applies GMM directly to linear Euler equations (see
Pindick and Rotemberg, 1983; Fair, 1989; Eichenbaum, 1990), Hansen and
Sargent's method described above requires more assumptions about the stochastic
law of motion of economic variables but utilizes more restrictions and is asymp-
totically more efficient when these assumptions are valid:
The unrestricted VAR approach will consistently estimate the system consist-
ing of (10) and (11) under the assumptions laid out above. It should be noted that
what the unrestricted VAR approach uncovers as the shock for P, is W" which is
given by (7). As evident from (7), W, is nothing but the difference between the
forecast of the present value with all information available to the economic agents
and the forecast of the present value with the econometrician's information set.
Thus in this example, the unrestricted VAR approach fails to identify shocks.
Conclusions
In this comment, a linear rational expectations model was presented, so that the
reader can compare the VAR approach with other competing methods. In the
example, the unrestricted VAR approach fails to uncover shocks. This example is
not meant to demonstrate that the VAR approach is inferior: of course, we can
construct other examples in which other competing methods fail to estimate para-
meters consistently while the unrestricted VAR approach is more successful.
However, the example emphasizes the fact that the disturbances in the VAR
approach contain the difference between the economic agents' forecast and the
linear forecast based on econometricians' information sets in general. Before in-
terpreting the shocks uncovered by the VAR approach, the econometrician needs
to think whether or not all important information available to the relevant eco-
nomic agents is included in the VAR system and whether or not economic agents
are using nonlinear forecasting rules in important ways.
Note
I. Maximum-likelihood estimation has been used more frequently for the Hansen-Sargent type
linear rational expectations model than GMM (see Sargent. 1978. 1981a. 1981b; Eichenbaum. 1984;
COMMENTARY ON CHAPTER 3 lOS
Finn, 1989; Giovannini and Rotemberg, 1989) even though Hansen and Sargent's (1982) method can
be applied to these models. Other related applications include West (1987, I988a).
References
Campbell, J.Y., and P. Perron. (1991). "Pitfalls and Opportunities: What Macroeconomists
Should Know About Unit Roots." In OJ. Blanchard and S. Fishcer (eds.), NBER
Macroeconomics Annual 1991. Cambridge, MA: MIT Press.
Campbell, J. Y., and R.I. Shiller. (1987). "Cointegration and Tests of Present Value Models."
Journal of Political Economy 95, 1062-1088.
Cooley, T.F., and M. Ogaki. (1995). "A Time Series Analysis of Real Wages, Consumption,
and Asset Returns Under Optimal Labor Contracting: A Cointegration-Euler Equation
Approach." Journal of Applied Econometrics, forthcoming.
Eichenbaum, M. (1984). "Rational Expectations and the Smoothing Properties of Inventories
of Finished Goods." Journal of Monetary Economics 14,71-96.
Eichenbaum, M. (1990). Some Empirical Evidence on the Production Level and Production
Cost Smoothing Models of Inventory Investment." American Economic Review 79,
853-864.
Engle, R.F., and WJ. Granger. (1987). "Co-Integration and Error Correction: Representa-
tion, Estimation, and Testing." Econometrica 55, 251-276.
Fair, R.C. (1989). "The Production-Smoothing Model Is Alive and Well." Journal of
Monetary Economics 24, 353-370.
Finn, M.G. (1989). "An Econometric Analysis of the Intertemporal General Equilibrium
Approach to Exchange Rate and Current Account Determination." Journal ofInterna-
tional Money and Finance 8,467-486.
Giovannini, A., and 1.1. Rotemberg. (1989). "Exchange-Rate Dynamics with Sticky Prices:
The Deutsche Mark, 1974-1982." Journal ofBusiness and Economic Statistics 7, 169-
178.
Hansen, Lars Peter. (1982). "Large Sample Properties of Generalized Method of Moments
Estimators." Econometrica 50 (July), 1029-1054.
Hansen, L.P., and TJ. Sargent. (1980). "Formulating and Estimating Dynamic Linear
Rational Expectations Models." Journal of Economic Dynamics and Control 2, 7-46.
Hansen, L.P., and TJ. Sargent. (1982). "Instrumental Variables Procedures for Estimating
Linear Rational Expectations Models." Journal of Monetary Economics 9, 263-296.
Hausman, JA. (1978). "Specification Tests in Econometrics." Econometrica 46 (Novem-
ber), 1251-1271.
Ogaki, M. (1993). "Generalized Method of Moments: Econometric Applications." Hand-
book of Statistics. Vol. 11, Econometrics.
Pindick, R.S., and U. Rotemberg. (1983). "Dynamic Factor Demands and the Effects of
Energy Price Shocks." American Economic Review 73, 1066-1079.
Sargent, T.J. (1978). "Rational Expectations, Econometric Exogeneity, and Consumption."
Journal of Political Economy 86, 673-700.
Sargent, T.J. (1981a). "The Demand for Money During Hyperinftations Under Rational
106 MACROECONOMET~CS
Expectations." In R.E. Lucas, Jr, and TJ. Sargent (eds.), Rational Expectations and
Econometric Practice. Minneapolis, MN: University of Minnesota Press.
Sargent, TJ. (1981b). "Estimation of Dynamic Labor Demand Schedules Under Rational
Expectations." In R.E. Lucas, Jr, and T.J. Sargent (eds.), Rational Expectations and
Econometric Practice. Minneapolis, MN: University of Minnesota Press.
West, K.D. (1987). "A Specification Test for Speculative Bubbles." Quarterly Journal of
Economics 102, 553-580.
West, K.D. (1988a). "Dividend Innovations and Stock Price Volatility." Econometrica 56,
37-62.
West, K.D. (1988b). "Asymptotic Normality, When Regressors Have a Unit Root."
Econometrica 56, 1397-1417.
West, K.D. (1989). "Estimation of Linear Rational Expectations Models in the Prescence
of Deterministic Terms." Journal of Monetary Economics 24, 437-442.
4 PROGRESSIVE MODELING OF
MACROECONOMIC TIME SERIES
The LSE Methodology
Grayham E. Mizan
Introduction
Econometric models, large and small, have played an increasingly important role
in macroeconomic forecasting and policy analysis. However, there is a wide range
of model types used for this purpose, including simultaneous-equation models in
either reduced or structural form, vector autoregressive models (VAR), autoregres-
sive distributed-lag models, autoregressive integrated moving-average models,
leading-indicator models, and error-correction models (ECM). Hendry, Pagan,
and Sargan (1984) discuss a typology for dynamic single-equation models for
time-series variables, and Hendry (1994) presents a typology for the various types
of dynamic model used in the analysis of systems of equations. There is also a
wide range of views about the appropriate way to develop and evaluate models.
Sims (1980, 1992) advocates the use of VAR models, which can accurately rep-
resent the time-series properties of data, while eschewing the reliance on "incred-
ible dentifying restrictions" that characterizes the use of simultaneous equation
models of the structural or Cowles Commission type. The potential value of struc-
ture (loosely defined) within the context of VAR models has led to the develop-
ment of structural VAR models, and Canova (1995) provides a recent review of
this literature. Leamer (1978, 1983), on the other hand, has been critical of the use
107
108 MACROECONOMETRICS
of non-Bayesian models that do not analyze formally the role and value of a priori
information, especially when there is no checking of model sensitivity. Summers
(1991), though aware of the important developments made in theoretical statistics
and econometrics in this century, argues that too much emphasis is placed on the
technical aspects of modeling and not enough on the real issues that are concerned
with the analysis of well-established and fundamental relationships between eco-
nomic variables. One approach to modeling that does not overemphasize the role
of model evaluation and statistical technique is that associated with real business
cycle analysis and the calibration of economic theory, rather than its evaluation.
Kydland and Prescott (1982, 1995) have been pioneers in this field, and Canova,
Finn, and Pagan (1994) provide a critique.
The focus of this paper is an alternative approach to the modeling of economic
time series that was originally developed at the London School of Economics
(LSE), though contributions and extensions to this methodology have subsequently
been made by econometricians who have had no direct connection with LSE
as an institution. The essence of this approach is the recognition that potentially
valuable information for the analysis of any economic problem can come from
numerous sources, including economic theory, the available sample of observa-
tions on the potentially relevant variables, knowledge of the economic history of
the period under study, and knowledge of the way in which the observed data are
defined and measured and their relationship to the theory variables. In the devel-
opment of econometric models it is therefore important that information from all
these sources is exploited as fully as possible. Of course, the marginal return to the
exploitation of information from each source will vary across the sources and the
degree to which each source has been used already. However, attention here will
be confined to consideration of the extent to which available information has been
exploited and to the relationship between this and the evaluation and comparison
of alternative models of the same phenomena.
The next section provides a brief history of the group of econometricians
involved in the development of the LSE modeling methodology. This is followed
by a section presenting the essential components of the LSE methodology, empha-
sizing the importance of evaluating and comparing altemative models within a
statistically and economically coherent framework. This approach helps to ensure
that econometric modeling is progressive by only abandoning models found to be
inadequate in favor of models that have been demonstrated to be improvements.
It is also progressive in the sense that it is not necessary to know the complete
structure characterizing the relationship between economic variables prior to
commencing an analysis of it. Rather, it is possible to discover incrementally parts
of the underlying structure as a result of careful analysis (see Hendry, 1993b).
The following section contains an analysis of the time-series relationship between
wages, prices, and unemployment in the United Kingdom between 1965 and 1993,
LSE METHODOLOGY 109
in this area. The weekly Econometrics Workshop, associated with research pro-
grams funded by the Social Science Research Council, was a focus for the devel-
opment and discussion of research ideas internally, but it also benefited from the
participation of longer-term visitors such as Takeshi Amemiya, Ted Anderson
(see his foreword to Maasoumi, 1988), Rob Engle, Alain Monfort, and Charles
Nelson. Interaction with other econometricians came via periods of leave spent by
members of the LSE econometrics group at the Australian National University,
University of California at Berkeley, CORE, Yale, and particularly at University
of California at San Diego. These visits provided valuable opportunities to prove
and enrich ideas, and many lengthy, vigorous, and sometimes provocative conver-
sations with Rob Engle, Clive Granger, Michel Mouchart, Adrian Pagan, Tom
Rothenberg, and Hal White stimulated further developments and resulted in some
joint papers.
By the end of the 1970s the LSE methodology was becoming more widely
known as a result of journal publications, presentation of papers to conferences
and in particular to the UK SSRC Econometrics Study Group, and via the ex-
panding group of former students such as Gordon Anderson, Richard Blundell,
Julia Campos, James Davidson, Neil Ericsson, Chris Gilbert, Mark Salmon, and
Aris Spanos. Also many of the estimation and testing procedures developed as an
integral part of the methodology were implemented in the computer program
written by David Hendry-GIVE (generalized instrumental variable estimation),
which was widely distributed and used. GIVE and other programs in the AUTOREG
library (see Hendry and Srba, 1980) were the mainframe precursors of the pro-
grams PcGive and PcFirnl (see Hendry, 1989; Doornik and Hendry, 1992, 1993,
1994) used in the empirical work reported in the fourth section below.
Also at the end of the 1970s some members of the LSE econometrics group
took up positions at other universities in the UK: Wallis went to Warwick, Hendry
to Oxford, and Mizon to Southampton. Following this and the later retirement of
Denis Sargan the methodology was much less clearly identified with the LSE.
Indeed, the LSE has via Andrew Harvey (Statistics Department) and Peter Robinson
(Economics Department) fostered other important aspects of econometrics, and
the subsequent development of the methodology, especially for the analysis of
systems of integrated-cointegrated variables, has been greatly influenced by the
contributions of Peter Phillips (Yale), Sf/Jren Johansen (Copenhagen), and Katarina
Juselius (Copenhagen) initially, and more recently by those of Peter Boswijk
(Amsterdam) and Jean-Pierre Urbain (Limburg). Further, the initial development
of the theory of encompassing by Hendry, Mizon, and Richard (see Hendry and
Richard, 1982, 1989; Mizon, 1984; Mizon and Richard, 1986) has been much
extended and refined by the contributions of Jean-Pierre Aorens (Toulouse), Neil
Ericsson (Federal Reserve, Washington), and Maozu Lu (Southampton) among
others. Gourieroux and Monfort (1995) provide a recent contribution. Hence there
112 MACROECONOMETIUCS
are many people who have contributed to the development of the LSE methodol-
ogy, and many more now ensuring that it is refined, improved, and applied. The
next section provides a brief description of its main components.
There have been many papers written on this topic (Ericsson, Campos, and Tran,
1990; Gilbert, 1986, 1989; Hendry, 1987; Hendry and MilOn, 1990; Hendry and
Richard, 1982, 1983; MilOn, 1991; Pagan, 1987, 1989; Spanos, 1986; and other
papers in Granger, 1990). Hendry and Wallis (1984), which contains contributions
by some of Sargan' s former colleagues and students to mark his retirement, also
provides an insight into the LSE methodology. Hence only a brief discussion of
the main components of the LSE methodology will be presented in this section,
together with comments on the nature of models and their relationship with the
data generation process (DGP).
By way of illustration consider an artificial situation in which it is known that
observations on y" x" and z, (which might be wages, prices, and unemployment,
respectively) are independent drawings from a trivariate normal distribution with
mean vector J.L and covariance matrix I. Hence the DGP for y" x" and z, is
independent normal with mean vector J.l and covariance matrix I with
yx
J.LY) (GYY G GYZ) (4.1)
Jl = J.Lx I = Gxy Gxx Gxz '
(
J.L z GZy Gzx Gzz
that is, the joint distribution of Yr' X" and z" denoted D(y" X" z,; J.L, 1:), has the form
D(y" X" z,; J.L, 1:) =Nl(J.L, 1:). However, this joint density can be reparameterized
as the product of conditional and marginal densities:
D(y" X" z,; J.L, 1:) = D(x, I z" y,; Ox, (011) X D(z, I y,; Oz, (022) X D(y,; 0., (033)
with
and qJ = O'yyO'u - 0';,. Adopting the (8, il) parameterization the DGP can be
described in the equations
(4.2)
Now, let the equations (4.2) be the equations used to generate the data in a Monte
Carlo simulation study, so that the investigator who writes down the model
(4.3)
with "It - NI(O, 0';1) is clearly seen to have chosen a "true" model since it corre-
sponds exactly to an equation of the DGP, with C 1 = C" PI = /3, Yt = y, and 0';. =
(QII. Indeed, M I corresponds to D(x, I z" y,; 8" (QII)' which is part of the DGP.
Hence provided that the phenomenon of interest to the investigator can be analyzed
solely in terms of the parameters of M I , and y, and z, are weakly exogenous
variables for the parameters of interest, no problems will arise for that investigator.
The fact that D(x, I z" y,; 8" (QII) is only part of the DGP so that M I is not
exploiting all the information in the DGP, is the reason for requiring y, and z, to
be weakly exogenous. If the parameters of interest to the investigator are denoted
¢, then y, and z, will be weakly exogenous for ¢ provided that (I) ¢ can be
recovered uniquely from 8, and (QII' and (2) the values of 8, and (QII in the DGP
are chosen independently of the values of the other parameters of the DGP-
namely, 8" 8y' COn and (Q33. The requirement (2) is satisfied for the DGP being
discussed here provided that there are no a priori restrictions linking (8;, (QII) to
(8;, 8y , (Q22' (Q33) (for example, a =/3). Engle, Hendry, and Richard (1983) provide
fuller definitions and discussion of this and other concepts of exogeneity.
Now consider another investigator who claims to be interested in the model
(4.4)
With "2, - NI(O, 0';2) this investigator is clearly not considering the "true" model
since M2 does not correspond to any of the equations of the DGP as given in (4.2)!
It is, though, part of other valid parameterizations of the DGP:
114 MACROECONOMETRlCS
D(y" x" z,; J.L, I) = D(YI I X" z,; ~y) X D(x, I z,; ~x) X D(z,; ~,)
= D(y, I X" Z,; ~y) x D(z, I X,; ',) x D(x,; 'x)
and thus "true". Further, the investigator who chooses to analyze the model
(4.5)
with U3, - NI(O, O"~) is clearly misguided in ignoring the influence of z, on Y, and
could be argued to be using a misspecified model, in addition to not having a
"true" model. However, M3 need be neither misspecified nor untrue. If the param-
eters of interest to the investigator are those of the distribution Y, I x, (that is, ~ in
D(y, I X,; ~», then M3 is obtained as a reduction of the DGP given by NI(J1, I)
and as such is both true and correctly specified when C3 = J.Ly - a 3 J.Lx and a 3 =
O";;O"xy with 0";3 = O"yy - a~O" xx' Alternatively, if M3 is thought to be a statement
about the distribution D(y, I X" z,; ~y), then it will only be fully exploiting the
available information if the conditional independence hypothesis y,.l z, I x, (or Y2
=0 in M 2) is valid.
Two important points have been illustrated in this discussion. First, all models
can be obtained as reductions of the DGP for the variables they involve, and
therefore their properties are implied by the DGP. In fact, the parameters of any
model are appropriately interpreted as the pseudo true values (derived under the
DGP) of the maximum-likelihood estimators given the model specification.
Second, it is important in all econometric studies to know what the parameters of
interest are or, equivalently, what the relevant statistical distribution and informa-
tion set are. Hence a model can be relevant only if it is defined with respect to an
appropriate information set and if the parameters of interest can be uniquely
derived from its parameters. But this is not sufficient for efficient inference on the
parameters of interest. It is also important that the model be congruent with the
available information. For example, a necessary condition for a model like M3 to
be congruent, when the parameters of interest are functions of ~Y' is Y2 =O. This
is a testable hypothesis, and it is important that it be tested as a part of the
evaluation of the model.
The importance of testing and evaluating econometric models is central to the
LSE methodology. This was clear in Sargan (1964), in which tests for functional
form (natural logarithm In versus level), for appropriate dynamic specification (the
order of dynamics and common factor restrictions (COMFAC», and for instru-
ment validity were proposed and used. Subsequent papers that further developed
this theme include Hendry and Mizon (1978), Hendry and Richard (1982, 1983),
Mizon (1977a), and Hendry (1980), this last paper concluding with the recommen-
dation to "test, test, and test." Hypothesis testing can be used both to assess the
adequacy of the currently specified model (misspecification testing) and to explore
the possibility for simplifying the current model without losing coherence with
LSE METHODOLOGY 115
available infonnation including the observed data (specification testing) (see Miron,
1977b, for discussion of this distinction). Both these fonns of testing play impor-
tant roles in the evaluation of econometric models, an issue to which attention is
now turned.
If an econometric model is to be taken and used seriously, then its credentials must
be established. Two important ways to do this are to demonstrate that the model
is coherent with the available relevant infonnation and that it is at least as good
as alternative models of the same phenomena. Hence congruence and encompass-
ing are central to the LSE methodology.
A Priori Theory. That for econometric models the primary source for a priori
theory is economics and economic theory is obvious and uncontentious. The rela-
tive importance of economic theory and statistical criteria in the development and
evaluation of econometric models, though, has been the subject of much debate,
with probably the most frequently cited example being the "measurement without
theory" debate between Koopmans and Vining (see Hendry and Morgan, 1995, for
a recent assessment of this debate). However, the tensions between VAR modeling
and real business cycle and calibrated models are a current manifestation of the
debate that still surrounds this issue. In crude tenns the contrast is between data-
driven modeling in the use of VAR's, and theory-driven modeling of either the
structural type (a la Hansen and Sargent) or the calibrated real business cycle type.
Canova (1995) presents a survey of VAR modeling, Ingram (1995) reviews struc-
tural modeling, Kydland and Prescott (1995) present the case for calibration
methods applied to models of the real business cycle, and Kim and Pagan (1994)
discuss ways in which this latter class of model can be evaluated. In this context
the LSE approach draws on the strengths of each of these alternative approaches,
though this is not the way that the LSE methodology evolved. Hence Hendry
(1993b) argues that while it is desirable that econometric models be consistent
with an economic theory, it is not essential. Indeed, if all models were required to
have their origins in existing economic theory, then the scope for developing new
theories and learning from observation would be greatly reduced. Further, there is
nothing that endows an economic theory with veracity Cl priori, and so coherence
116 MACROECONOMET~CS
information is {XI' X2' ... , X,-I}, and the relative future information is {X'+I' X'+2'
xr} when T is the final observation available in the sample.
Xt+3' ••• ,
A model that is not congruent with past sample information will have errors
that are correlated with lagged values of X" and hence the errors will be serially
correlated and at least partially explainable in terms of {XI' x2, ... , X'_I}. For
example, if the OLS estimator of 6 in the static regression model,
Z, = AZI- I + V, (4.8)
when each of {£,} and {v,} are identically independently distributed processes that
are independent of each other with I a I < I, I AI < I, so that both y, and z, are
stationary. Then in the static long run (that is, when y, = y* and z, = z* 'Vt) the
response of y, to z, is given by 1(= (13 + y)/(I - a), whereas the pseudo true value
of the OLS estimator of 6 is given by 4 =(13 + yA)/O - aA). Hence inferences
about the long-run response of y, to z, based on the static regression model (4.6)
will be invalid unless I( = 60, Note that I( = 60 if and only if y= - af3, which is
the common factor restriction that ensures that the autoregressive distributed lag
model (4.7) takes the form
y, = f3z, + £"
£, = a£,_1 + E" (4.9)
2. Relative Present Sample Information: A model with errors that are not
homoscedastic innovations with respect to the set of all current dated variables in
the modeler's databank of relevant variables is not congruent and can be improved
by using information already available. For example, if a model's errors u, are
*
such that E(u,x,) 0 then there is still information in X, that is unexploited. This
could arise from erroneously omitting variables from the model or by conditioning
on variables that are not weakly exogeneous for the parameters of interest. An-
other form of this type of noncongruence is heteroscedasticity in the errors (that
is, E(u~) =(J'~* (J'2'<:/t), which implies that the distribution of the errors has features
that are potentially explainable as functions of the X" The distribution of the errors
may also be skewed, leptokurtic, or platykurtic, and to the extent that these fea-
tures of the error distribution are explainable as functions of the available data the
model is noncongruent. Tests for omitted variables, for weak exogeneity (see
Engle, Hendry, and Richard, 1983; Ericsson, 1993; Richard, 1980), excess skewness
or kurtosis in the error distribution (see Doomik and Hansen, 1994), and
homoscedasticity (see Breusch and Pagan, 1980; White, 1980) are examples of
ways in which a model's congruence with present sample information can be
assessed.
itself ephemeral, and inferences based on it are likely to appear whimsical when
judged against subsequent observation. Indeed, such a model could not be captur-
ing underlying structure. A more durable model will have its parameter estimates
approximately constant across varying estimation periods. Further, when a model
is to be used for prediction of the likely effects of a change in policy (such as a
change in tax rates, a move from fixed to floating exchange rates, or the intro-
duction of a tight monetary policy), it is crucial that the model's parameters are
invariant to the policy regime shift. Discussion of this issue has a long history,
early contributors being Frisch and Haavelmo who in their analysis introduced the
concept of autonomy (see Aldrich, 1989). Lucas (1976) presented a more recent
view of the problem, which has spawned a large literature. Indeed, for many
macroeconometricians the Lucas critique was sufficiently pernicious to have re-
moved the credibility of much econometric modelling. However, if a model has
conditioning variables that are weakly exogenous for the parameters of interest,
and the latter are invariant to changes in the process generating the conditioning
variables (the conditions for superexogeneity; see Engle, Hendry, and Richard,
1983), then the Lucas critique is rendered benign. An important implication of this
statement is that the Lucas critique is brought into the realms of testable hypo-
theses as explained by Engle and Hendry (1993) and Favero and Hendry (1992)
(see Ericsson, 1993, for a recent appraisal). The main instruments of such testing
are analysis of variance type parameter constancy and Chow (1960) prediction test
statistics, as well as recursive estimation (see Brown, Durbin, and Evans, 1975;
Hansen, 1992; and Terasvirta, 1970).
rather than contentious, though the nature of the particular problem being analyzed
will determine the price of imprudence.
Rival Models. The importance of ensuring that a model is congruent with the
information contained in rival models is twofold. First, in economics there is
usually no shortage of alternative theories for a given phenomenon. Second, it is
possible for more than one model to be congruent with a priori theory, sample
information, and the properties of the measurement system. The fact that econ-
omics is a rich source of theories and hypotheses is a tribute to the economics
profession's ingenuity and imagination and a problem only if there is no serious
attempt to discriminate between the competing models. Typically, alternative
theories when implemented in econometric models use different information sets
and possibly different functional forms and are thus separate or nonnested models.
It is this nonnested feature that enables more than one model to be congruent with
respect to sample information: each can be congruent with respect to its own
information set. Ericsson and Hendry (1989) analyze this issue and show that the
corroboration of more than one model can imply the inadequacy of each, and
Mizon (1989) provides an illustration. Hence to have a model that is congruent
with respect to its own information set is a necessary condition for it to be exploit-
ing that information, but it is not sufficient for it to be a dominant or encompassing
model. An encompassing model is one which can account for the previous em-
pirical findings that were thought to be relevant and adequate for the explanation
of the variables and parameters of interest and can explain the features of rival
models being considered currently. The encompassing model renders other models
inferentially redundant and so is a dominant model.
The comparison of alternative models for the same phenomenon can be achieved
by using one of the many statistics for testing nonnested hypotheses (see Cox,
1961,1962; Davidson and MacKinnon, 1981; and Pesaran, 1974), but it is impor-
tant to note that the statistical (size and power) properties of these test statistics are
derived in the context of an embedding model-a point made clear by the encom-
passing interpretation of these test statistics (see Hendry and Richard, 1989; Mizon,
1984; Mizon and Richard, 1986). Further, the use of nonnested test statistics in
pairwise comparison of models can lead to apparent contradictions and is a
nontransitive procedure. In particular, it is possible for the hypotheses M,EM2 and
M2EM3 to be valid while the hypothesis M,EM3 is invalid. Since this nontransitivity
of encompassing arises essentially because the embedding model implicit in each
of the three pairwise comparisons is different (often M. U M2, M2 U M3 and M 1
U M 3, respectively), the problem can be avoided by comparing all models relative
to a general model that embeds or nests all of them. Letting Me denote such an
embedding model (that is, M; C Me Vi), it is then possible to test the nested
hypothesis that M; is an acceptable simplification of Me for each i. If there is a
LSE METHODOLOGY 121
model (M" say) that is a valid simplification of Me' then it parsimoniously encom-
passes Me-denoted M,EpMe. In the population it is possible for more than one
model to parsimoniously encompass an embedding or completing model Me only
if they are observationally equivalent, consequently, if MjEpMe, then M,E~'V j.
Therefore the requirement that a model be congruent with information contained
in rival models is equivalent to that model parsimoniously encompassing a model
that embeds all the rival models. This implies that the relevant sample information
set for the determination of a congruent model is the union of the information sets
of the models implementing each competing theory. Unsurprisingly, this entails
that if empirical evidence is to be used to evaluate and compare models rather than
naively corroborate them, a modeling strategy that starts from a congruent general
model and tests for valid simplifications or reductions of it is likely to be an
efficient way to find parsimonious yet durable representations of the salient fea-
tures of the relationships between the variables of interest.
Finally, in this section on the evaluation of model congruence, note that testing
for congruence with respect to sample information is misspecification testing,
whereas testing for congruence with respect to rival model information is speci-
fication testing. In this context misspecification testing can be seen as a means of
ensuring that there is a valid statistical basis for using specification testing to find
a parsimonious encompassing model. Provided that the sample information set
used ensures that all econometric models implementing relevant economic theo-
ries are nested within a general model that is congruent with the sample informa-
tion (this is determined by misspecification testing or diagnostic checking of the
general model), the model that parsimoniously encompasses the general model
and is data admissible will be the dominant model that is congruent with informa-
tion from all four sources (this is determined by specification testing or checking
the validity of reductions from the congruent general model).
is general enough to enable the empirical implementation of all the models. Hence
each model is evaluated with respect to an information set more general than the
minimum one required for its own implementation, thus achieving robustness to
extensions of its information set in directions relevant for competing models.
Modeling Strategies
The previous section discussed the evaluation of econometric models and em-
phasized the important roles of congruence and encompassing in this assessment.
Indeed, the most important features of a model are its quality and suitability for
its intended use, features that are assessed in the process of model evaluation and
not endowed on it as a result of the process by which the model is developed. The
route by which a model is discovered (including serendipity, brilliant intuition,
and systematic application of a particular modeling strategy) does not determine
model quality, though it does affect research efficiency) (see Hendry and Mizon,
1990, for further discussion of these points). Although a model with a good pedi-
gree (for example, derived from high-quality economic theory by a leading prac-
titioner) is more likely to be valuable than one without, such a pedigree is neither
necessary nor sufficient for the discovery of a congruent and encompassing model.
This is fortunate since otherwise there would be little or no scope for the discovery
of new theories and classes of model.
While there is no unique way to find congruent and encompassing models,
some modeling strategies are more likely to do so, and to do so efficiently, than
others. Of the many strategies that might be adopted in modeling, attention here
is confined to a contrast between specific-to-general and general-to-specific
modeling. Granger (1990) contains discussions of other possibilities.
(4.11)
with
(4.12)
This example, and others presented in Hendry (1993a) in the context of a bivariate
cointegrated system, illustrate the crucial role of weak exogeneity in sustaining
valid single-equation inference with cointegrated processes, despite the fact that
the OLS estimator is superconsistent in that context. This example indicates that
whether or not variables are weakly exogenous for the parameters of interest
depends on the properties of the joint distribution of all variables. Hence it is not
surprising that members of the LSE also paid attention to the issues of model
congruence and encompassing, and the desirability of using a general-to-specific
modelling strategy, in a systems context.
Although the importance of carefully considering the dynamic specification of
single-equation econometric models had been a crucial part of the LSE method-
ology since Sargan (1964), its relevance for systems of equations was also real-
ized. Discussions of dynamic specification in systems is contained in Hendry
LSE METHODOLOGY 125
(1971, 1974), Hendry and Anderson (1977), and Mizon (1977a). However, the
challenge to the econometric modeling of macroeconomic time series provided by
the work of Box and Jenkins (1970) and Granger and Newbold (1974) stimulated
other contributions. In particular, Prothero and Wallis (1976) analyzed the univariate
implications of systems of simultaneous equations, drawing attention to the par-
ticular form taken by the implied autoregressive integrated moving average
(ARIMA) time-series models (also see Zellner and Palm, 1974). These and other
contributions, for which see the discussion in Hendry, Pagan, and Sargan (1984),
were concerned with the specification, estimation, and testing of multivariate
econometric models with respect to their own information sets and did not raise
the issue of one structural econometric model encompassing its rivals. Noting that
the typical specification of a dynamic linear structural econometric model (SEM),
based on the pioneering work of the Cowles Commission researchers (see
Koopmans, 1950), has the generic form
when Y, are the endogenous variables being modeled, z, are weakly exogenous
variables that are not modeled, X;_I = (Y:-J' Z:_I) are the lagged values of both
endogenous and exogenous variables, and u, is the vector of unobserved errors, it
is clear that the underlying distribution is D(Y,I z" Xt-I), which corresponds to the
unrestricted reduced form that in conventional notation can be written
which has proved to be a valuable way to model the joint distribution of x, con-
ditional on its history D(x, I Xo, XI> ••. ,Xt-I). Monfort and Rabemananjara (1990)
for stationary X, illustrated the use of a VAR as a general model within which to
test the exogeneity status of a subset of variables and to test particular structural
hypotheses. Hendry and Mizon (1993) proposed the use of a congruent VAR when
x, is nonstationary (x, - I(l) and/or the process-generating x, has deterministic
nonstationarities such as trends and regime shifts) as a general statistical frame-
work within which to test hypotheses about the dimension of cointegrating space
following Johansen (1988), as well as hypotheses concerning dynamic simplifica-
tion (such as Granger noncausality), weak exogeneity, and the ability of particular
structural econometric models to parsimoniously encompass the VAR. Note that
in this framework the evaluation of each SEM is relative to the infonnation set
supporting the VAR and not just the SEM's own infonnation set. Clements and
Mizon (1991) and Hendry and Doornik (1994) contain applications of the general-
'to-specific modeling strategy for I(l) systems proposed by Hendry and Mizon
(1993), and the next section provides an illustration using an updated databank for
a subset of the variables modeled in Clements and Mizon (1991).
Just as the importance of evaluating models by checking their congruence and
encompassing abilities is not confined to single-equation models, it is not limited
to models for time-series variables. It is just as important for cross-section and
panel data models that their congruence be tested and that the encompassing
properties of alternative models are assessed, rather than simply using empirical
analysis to calibrate and confinn particular economic theories. Research devoted
to furthering the development of model evaluation criteria and test statistics for
microeconometric models, analogous to those developed for macroeconomic time
series models, could yield high returns in tenns of deepening and putting on a
finner basis our understanding of microeconomic relationships. Since general
tools for estimation and hypothesis testing are available, the need is for a
LSE METHODOLOGY 127
_~k2::1 :-:(21
1975
log real wages
1985 1995 2.4 1975
3.2...-..:.;log:2...:.::%-:u::..:n=e::':"L:.:I:=llI=en:.:.;t:.....:..r=at;:,;e::....-_--,
~'rmm::J ~'1IillJTIJ
.8 .8
.7 .7
.6 .6
.5 .5
.4 .4
.3 .3
.2 .2
.1 .1
Ii! II
5 1 Ii! 15 5 1 Ii! 15
As an illustration of the use of univariate test statistics for units roots, Table 4.1
provides the values of augmented Dickey-Fuller test statistics, which take the form
of the r statistic for the hypothesis a == 0 in the regression model for the generic
variable Y,:
w -1.27 2
u -1.77 1
p -1.31 1
(w - p) -1.10 3
<iw -3.30* 1
<ip -3.83** o
du -4.95** o
* Significant at 5 percent.
** Significant at I percent.
seasonal dummy variables were included in the regression (4.17). On the basis of
these augmented Dickey-Fuller statistics all three variables w, p, and u appear to
be I(I)-the null hypothesis of 1(0) being rejected against the alternative of 1(1)
and the null of 1(1) not being rejected against the alternative of 1(2) for each
variable. In addition, the real wage (w - p) is 1(1), and so w and p do not cointegrate
as the real wage. In fact, the hypothesis that each of these variables has a unit root
was not rejected when tested against the alternative of trend stationarity-that is,
a; 0 was not rejected even when a linear trend was added to the regression (4.17).
Hence in the modeling of w, p, and u it will be important to choose models that
can represent their nonstationarity, with the possibility that all three variables fonn
a cointegrating relationship-that is, there is some linear combination of w, p, and
u that is 1(0). Another characteristic of wages and prices is that they are nonnegative,
so the use of linear models in the logarithmic transfonnations of them is data
admissible-that is, cannot produce negative fitted or predicted values. Similarly,
the unemployment rate is bounded between zero and 100 percent, so that a logit
transfonnation might be appropriate for modeling it in a data admissible way.
However, since all observations on the unemployment rate are below 13 percent,
the use of its logarithm should be data admissible, and this is borne out by noting
that the graphs of u and the logit transfonnation of U are essentially identical when
adjusted for their different means.
Having briefly analyzed the univariate statistical properties of the variables to
be analyzed, attention is now turned to a review of the Phillips curve literature in
so far as it is relevant to the subsequent analysis.
Single-Equation Analysis
is in order. First, given that the primary purpose of the empirical analysis is to
illustrate the LSE methodology and that the methodology has important implica-
tions for system and single-equation modeling it is relevant to illustrate both.
Second, there are pedagogical benefits to presenting the simpler case before the
general, and the intention is to exploit these.
Many past and present members of the LSE Economics Department have ana-
lyzed models of wage and price determination (Desai, 1975, 1984; Espasa, 1975;
Layard and Nickell, 1985; Lipsey, 1960; Nickell, 1984; Phillips, 1958; Sargan,
1964, 1980a; Vernon, 1970), and so it seems particularly appropriate to present the
results of re-analysis of this relationship with the present dataset. The first and
most influential piece of research in this area was that of A.W.H. Phillips, in which
an empirical relationship representing a tradeoff between money wages and unem-
ployment was described. As Desai (1984, p. 253) points out the Phillips curve "has
been re-specified, questioned, rejected as being unstable and reinterpreted," and
yet it is still evident in many models of aggregate labor market and wage and price
behavior. The Colston paper of Sargan (1964), though primarily concerned with
a number of methodological issues that had a major influence on LSE econo-
metricians (see Hendry and Wallis, 1984), developed single-equation models of
wage and price determination with the distinction between equilibrium and
disequilibrium behavior explicit. Indeed, the foundations of the error-correction
model, which is now so prevalent in the analysis of nonstationary-integrated data,
were laid in this paper. An additional feature of Sargan's paper was the extension
of the Phillips framework to allow for the impact of productivity. The results
presented below, however, are confined to describing relationships between wages,
prices, and unemployment. This is done in order to keep the empirical modeling
as simple as possible, while still providing an illustration of many of the important
points made in the previous discussion of the LSE methodology. In order to
increase the economic and statistical credibility, as well as the applicability, of the
model, it may well be necessary to enlarge the system to include productivity,
hours of work, exchange rates, and interest rates. In addition, it may be important
to pay detailed attention to the changes that have taken place in the U.K. labor
market since 1979, such as the increase in part-time working, the increase in self-
employed labor relative to employees, and the increased participation of women
in the labor force. Thus the class of model considered in this section can be written
in error correction form as
when k is the maximum lag and with the normalization a lO = I. A linear trend
is included in (4.18) as a proxy for omitted variables such as average labor
132 MACROECONOMETRlCS
Note that the defining characteristics of the static equilibrium are zero growth rates
(~w* =~p* =~u* =0) and no autonomous trend (A, =0). When the growth rates
are not zero, equation (4.19) can be rewritten as
E(w, I P" u,) = (Po + PIP, + fJIu,) + A,*t + (a'fA,* - ~)
Diagnostic statistics:
R 2 = 0.768 (1= 0.0099 v= 0.325
J = 2.581 AR(5, 82) = 0.712 ARCH(4, 79) = 0.642
[p> 0.62) [p > 0.63)
N= 0.042 H(40, 46) =0.756 RO, 86) =0.851
[p> 0.98) [p> 0.82) [p > 0.36)
. 027.-----------------, 1.2.-----------------,
..............................................
• 018
.9
.009
o .6
-.009
-.018 ........................- .
.8
.8
.6
.6
.4
.4
.2
.2 0
1980 1985 1990 1995 19811 1985 19911 1995
residuals, the residual correlogram, and the residual frequency plot and density.
Further, re-estimation of the model by recursive least squares, starting from an
initial sample of 50 observations and sequentially increasing the sample from 50
to 108, revealed no serious indications of overall nonconstancy. The graphs of the
recursively computed one-step ahead residuals bordered by ± twice their standard
errors, and the one-step ahead Chow, breakpoint-F and forecast-F test statistics in
Figure 4.4 confirm this, though there is a suggestion of some nonconstancy around
1984, which was the year of the miners' strike. Doornik and Hendry (1992, 1994)
provide the definitions of these statistics, as well as discussing their role in modeling.
However, the recursively computed estimates of the constant and the coefficients
of W,_I, P,-I, and U'_l show some evidence of change, starting in 1979 and stabiliz-
ing at new values after 1984.
Hence this general model appears to be reasonably congruent with the available
information and so forms a valid basis for testing for further simplifications and
testing other hypotheses. The real-wage hypothesis /31 = I is rejected at 1 percent
significance since the test statistic for the hypothesis that the coefficients of W,_I
and PI-I sum to zero is F(I, 87) =5.00 [p> 0.028]. However, the hypothesis that
136 MACROECONOMET~CS
Diagnostic statistics:
R 2 = 0.753 (J = 0.0097 v= 0.299
J = 1.960 AR(5. 93) = 0.965 ARCH(4. 90) =
1.152
[p> 0.44] [p > 0.34]
N=0.228 H(20. 77) = 1.345 R(I. 97) = 0.621
[p > 0.89) [p > 0.18] [p > 0.43]
.9
"
-.9
-1.8
-2.'1
19'15 1985 1995 19'15 1985 1995
1.---------------,
.8
Residual Correlogrlllll
.2
Residulll Frequency Plot
.6 .16
.4
.2 I .12
" r-.....,Ir-'--'---'~.,.I.........-'--Ir-'----1
-.2
-.4
••
-.6 .IM
-.8
-1 l.-o.~-.........J5L..-..~-~1,...."'-'-~-'-:'
15 " -2
" 2 4
Diagnostic Statistics:
R 2 = 0.753 (1 = 0.0096 V=0.298
J = 1.680 AR(5. 94) = 0.941 ARCH(4, 91) 1.161 =
[p> 0.46] [p > 0.33]
N=0.233 H(18. 80) = 1.347 R(l, 98) =0.608
[p > 0.89] [p> 0.18] [p > 0.44]
138 MACROECONOMETRICS
eliminates u (but not ~) from the model. Not surprisingly, this remains a con-
gruent model, and the hypothesis that it parsimoniously encompasses the general
model of Table 4.2 is not rejected since the relevant statistic is F( II, 87) = 0.462
[p > 0.92]. The long-run solution (with the dummy variables D745 and Policy,
which were included in the estimation, omitted) of this final single-equation model
takes the form
ecm, = W, + 8.573 - 0.852PI - 0.008t.
(0.11) (0.05) (0.00 I) (4.24)
This equilibrium-correction mechanism is essentially the same as that estimated
from the general model reported in Table 4.2 and has the same interpretation. In
particular, the real-wage hypothesis is rejected, and in the longer term the level of
the unemployment variable has no effect on wages. This latter result is surprising
and will be investigated further in the following section. The estimated steady
state corresponding to equation (4.22) for this model has the form
E(w, I PI' u,) =- 8.61 + 0.85p, + 0.008t - 0.68b.p* + 0.14~*. (4.25)
By construction this steady-state equation is consistent with the growth of wages
implied by the long-run equation (4.24). An implication ofthis steady-state equa-
tion for w is that the equilibrium value of nominal wages is smaller (larger) ceteris
paribus the larger is the steady-state inflation rate when the latter is positive
(negative). Since the rate of inflation is almost always positive, this finding is
consistent with the view that higher rates of positive inflation are undesirable. The
other implication of the steady-state equation is that the equilibrium value of
nominal wages is larger (smaller) the larger is the growth rate of the percentage
unemployment rate when it is positive (negative). Hence the equilibrium value of
the nominal wage ceteris paribus is associated positively with increases in the
unemployment rate.
Given the simplicity of this illustration there are not many sophisticated eco-
nomic models that can be obtained from this data set and considered as alternative
explanations of the determination of nominal wages. However, in the spirit of
illustration the following models can used as examples of alternatives correspond-
ing to four of the commonly used single equation model types:
Mw • b.w, = c. + f3.b.p, + YI(Wt_. - /(IP,-I) + VII
MW2 b.w, = C2 + Y2(Wt-l - /(2UH) + V2,
MW3 (w, - P,) = C3 + A4 + V3'
(4.26)
These models are: M w1 , a first-order error correction model with wages adjusting
towards a static equilibrium that depends on prices only (that is, equation (4.18)
LSE METHODOLOGY 139
F(5, 98) = 9.15 F(6, 98) = 11.12 F(7, 98) = 10184.0 F(lO, 98) =29.86
Given that each of the models in (4.26) is nested within the congruent model
reported in Table 4.3 (this model rather than the final model of Table 4.4 is used
in order to include ",_I in the analysis), it is possible to test the validity of the
implied reductions from that model, and the corresponding parsimonious encom-
passing test statistics are reported in Table 4.6. Hence each of the four models fails
to parsimoniously encompass the general model (all the test statistics have p
values of zero), and so they are each inadequate characterizations of the relation-
ship between w" PI' and ",.
At this stage the model that performs most satisfactorily is that reported in
Table 4.4, and so before turning to system analysis of the relationships between
w" P" and ", the forecasting ability of this model will be evaluated. Firstly, this
model was re-estimated with sample data for 1966(1) to 1989(3) and then used to
forecast aw from 1989(4) to 1993(1). The outcome is reasonable as indicated by
the graphs in Figure 4.6. In addition, the X2 prediction test statistic at r(l4) =
23.01 [p > 0.06] and the Chow (1960) prediction test statistic at F(14, 85) = 1.35
[p> 0.20] do not reject the null of parameter constancy, though this result is very
marginal for the X2 prediction test statistic. Indeed, Lu and Mizon (1991) showed
r
that the implicit null hypothesis of the prediction statistic is (x;atN + (CT~ - CT~)
=0 "if t e [(T + I), (T + H) J, and that of the Chow statistic is known to be X 2M3
= 0, when for the generic linear regression model y, = x;f3 + ", the data for the
sample period (t = 1,2, , T, which is denoted by subscript 1) and prediction
period (t =(T + 1), (T + 2), , (T + H), which is denoted by subscript 2) can be
written as
Hence the X2 prediction test statistic is sensitive to both changes in the error
variance and in the regression coefficients via changes in the conditional mean of
y" whereas the Chow prediction statistic is powerful against changes in the con-
ditional mean only: a constant error variance (a~ = a~) is part of the maintained
LSE METHODOLOGY 141
Du Actual a f it ted
.J.5,........;;.::...~:..::::.;;....::....:....;.::...=..::...---.---..., 3
Dw Residuals
.J.2 2
.89 J.
-J.
-2
-3
J.":I 197:1 J.'8:1 J.,,:I
•82
,.
I
, .88
I
I
8 .84
-.82 8
J.'98 J.'95 -2 8 2 4
hypothesis for the Chow statistic. Clearly, though, if the error variance is not
constant and/or the error distribution is otherwise nonstationary, the Chow statistic
will not have a central F(H, T + H - k) distribution under the hypothesis X2t1f3 =
O. Bearing in mind that 1989(4) saw the tightening of monetary policy following
the United Kingdom joining the European Monetary Mechanism, the finding that
there might have been a change in the model's error variance is not surprising,
even though it is a clear limitation of the model. However, when secondly the
model was re-estimated with data for the period 1966( I) to 1979(2), and then used
to forecast t1w over the period 1979(3) to 1993(1) the result was much less satis-
factory. The x:
prediction test statistic over this period takes the value X2(55) =
326.2 [p > 0.00], which strongly indicates parameter nonconstancy in the model.
The Chow prediction statistic on the other hand takes the value F(55, 44) =0.962
[p > 0.56] and so still does not reject the null of parameter constancy-but there
is evidence that the error does not have a stationary distribution throughout the
period 1966(1) to 1993(1). Inspection of the graphs in Figure 4.7 does reveal the
poor forecasting performance of the model for the period 1979(3) to 1993(1).
The estimates for the final model over the sample period 1966( I) to 1979(2) are
given in Table 4.7. and these are noticeably different from those for the full sample
142 MACROECONOMETRlCS
....t __
.15 ,.......:DIlI:..=.....;Ac~t..:.:ua.:;;I=--=It_r-,i ::d
t\__ --., Dw Residuals
4..--=-=-====-,.--------,
.12
.12
. .,
I
.113 .88
.84
Diagnostic statistics:
R 2 = 0.823 a = 0.0097 v= 0.176
J = 0.759 AR(5. 39) 2.599* = ARCH(4. 36) = 0.377
[p > 0.04] [p > 0.82]
N= 0.930 H(17. 26) = 1.210 R(1. 43) = 0.107
[p > 0.63] [p > 0.32] [p> 0.74]
LSE METHODOLOGY 143
1966(2)-1979(2)
and a rapid increase in unemployment. Both these features are clearly evident in
Figure 4.1.
Therefore there appears to be evidence that the single-equation models de-
veloped above for the explanation of w, in terms of PI' UI , and t have not captured
a constant underlying structure. Note that this has arisen despite the fact that the
full sample estimated models appeared to be congruent, with neither of the Hansen
(1992) instability statistics V and J indicating any structural break around 1979-
1980. Possible reasons for this include lack of cointegration between w" P" and u,
and the invalidity of the assumption that P, and u, are weakly exogeneous vari-
ables. These issues are related and best analyzed in the context of system modeling,
to which attention is now turned.
System Analysis
(4.30)
when x; =(WI' PI' u,), D I contains deterministic variables such as a constant, trend,
seasonal, and other dummy variables, and £, is a three-dimensional error vector
that is independently distributed with mean zero and covariance matrix I. It is
assumed that the roots of det (I - L~.AjLj) =0 lie on or outside the unit circle so
that there are no explosive roots, and that k is finite so that moving average-error
processes do not fall into the category of model being considered. The initial
conditions X I_ h X 2- k' ••• , X o are assumed fixed, and the parameters (AI' ... , A k ,
lP, I) are required to be constant in order for the adopted inferential procedures
to be valid. An observationally equivalent parameterization of (4.30) is provided
by the following vector equilibrium correction model (VECM):
(4.31)
n
in which = (I - LJ=IA) is the static equilibrium response matrix, and j = (1- n
Lk.Aj) for j = I, 2, ... , (k - 1) are the interim multiplier matrices. In the case that
n
'" - 1(1) the rank of is determined by the number of cointegrating vectors.
LSE METHODOLOGY 145
---J
Dp Actual a Fitted Dp residual
~!~,
------1 :~~-1
1975 1985 1995 1975 1985 1995
price controls; Budget, which takes the values I and -I respectively in the second
and third quarters of every year and zero in all other quarters and represents fixed
seasonal effects and especially the changes in excise duties and other tax rates,
announced in the annual budget, on prices and the rate of inflation; and Expansion,
which takes the value 1 in 1966(3) -1 in 1967(2), 1972(3), and 1974(3), highlight-
ing quarters in which fiscal policy was used to stimulate(+)/dampen(-) the economy,
and takes the values 2 in 1971(2) -2 in 1971(1) to represents the effect of the
announced cut of 50 percent in Selective Employment Tax. Since these four
dummy variables should not have a long-run effect on any of the modeled vari-
ables w" P" and U they are entered unrestrictedly into the VAR when the reduced
"
rank maximum-likelihood procedure of Johansen (1988) is applied it. However,
there is no evidence of quadratic trend in any of the modeled variables and so the
linear deterministic trend is restricted to lie in the cointegration space.
When an unrestricted VAR parameterized as (4.31) with k = 4, a constant and
the four dummy variables plus a linear trend entered as described above, was
estimated over the period 1966(1) to 1993(1) it produced a plausible and more-or-
less congruent system. Figure 4.8, which gives the graphs of the actual and fitted
LSE METHODOLOGY 147
* Significant at
values for 8W,. 8p" and !w." and the corresponding residuals (scaled), illustrates
this point. The system provides a high degree of explanation for the changes 8w,.
8p" and 8U, in the three modeled variables WI' p" and u, (it is. of course, even more
impressive in terms of the levels of the variables), and there does not appear to be
any strong heteroscedasticity or serial dependence in the residuals. Table 4.9
provides descriptive statistics for the estimated system: first single-equation re-
sidual standard deviations 0', serial correlationAR. autoregressive heteroscedasticity
ARCH. heteroscedasticity H, and normality N test statistics as defined for Table
4.2, and second test statistics for vector autoregressive residuals veeAR, vector
heteroscedasticity veeH, and finally vector normality veeN (see Doornik and
Hendry, 1994, for more details of these test statistics). The only evidence of non-
congruence comes from the AR test statistic for the t¥, equation and the veeAR
statistic. both of which are significant at 1 percent. Further evidence on the ap-
proximate congruence of the system is provided by the residual correlograms and
frequency plots given in Figure 4.9.
Though there is some evidence of residual serial correlation in the equations for
8W, and 8p" it is not strong. and experimentation revealed that inappropriate lag
length k does not appear to be the cause of this problem. In fact, inspection of the
residual plots in Figure 4.8 suggests that the difficulty may still lie in the system's
inability to represent the many changes in government policies towards wages and
prices, interest rates and exchange rates, and unemployment during the sample
period. However. rather than introduce more event specific dummy variables into
the analysis, which would increase the risk of the system being too finely tuned
to the particular sample data and hence even more likely to suffer from predictive
failure, it was decided to continue with the present set of dummy variables. The
other alternative of extending the information set to include interest and exchange
148 MACROECONOMETIUCS
:Iic,=.::=----j
_~ I ' , _ 5 ~GI ~5
~!:t"---L~-,-~·-~~~--1.:f ~
5 IG1 ~5 -2 GI 2 4
I
• 2~ ,---=-=-...:.:.::==c::....::..::.::c=.:=z...::..=;.:;"...---,
.~4
.Gl7
rates is not pursued here since the present modeling is intended to be illustrative
rather than definitive, but it is the subject of further research.
Inspection of the residual correlations in Table 4.10 suggests that there is a
modest correlation between W, and P" but the correlations between W, and U, and
P, and U, are negligible. The hypothesis that the order of lag required for each
variable is i can be tested using the statistics Fk=j, and on the basis of these it
appears that there is scope for simplifying the dynamics but that four lags are
required for P, consistent with it being seasonal. The estimated long-run matrix IT
has eigenvalues whose moduli are given by I Air I, and these indicate that there are
probably two zero eigenvalues suggesting that r = 1. This is confirmed by the
results of applying the Johansen (1988) maximum-likelihood procedure to esti-
mate the dimension of cointegrating space r, which are reported in Table 4.11.
The eigenvalues J.li' which are involved in the maximization of the log-
likelihood function with respect to f3 in order to obtain estimated cointegrating
vectors (see Johansen, 1988, or Banerjee et aI., 1993, for details), are small. How-
ever, the largest eigenvalue J.ll = 0.28 is significantly different from zero on the
basis of the maximum eigenvalue (Max = -Tlog(1 - J.lr» and trace (Trace =
- T~~=I 10g(1 - J.li» test statistics. r is the dimension of cointegrating space, and
LSE METHODOLOGY 149
w, P, UI
PI 0.14 1.0
U, 0.08 0.03 1.0
System Dynamics
w, PI U,
* Significant at
** Significant at
** Significant at
l (which is defined as - Tln:'~=, log( 1 - }J.;) ) is the corresponding value of the Iog-
likelihood function apart from a constant. Hence the value of 1445.87 for the
unrestricted log likelihood function is reported as 1446 for r = 3. The Max and
Trace test statistics are not adjusted for degrees of freedom as suggested by Reimers
(1992), since the results in KostiaI (1994) indicate a tendency for them to under-
estimate the dimension of cointegrating space even when unadjusted. The critical
values used for the Max and Trace statistics are given in Osterwald-Lenum (1992).
On the basis of the statistics in Table 4.11 it is concluded that there is one
cointegrating vector (r = 1), and an estimate of it is given by ~~ in Table 4.12.
This cointegration analysis with the linear trend t restricted to be in the
cointegration space yields an estimated cointegrating vector ~;x, = w, - 0.88p, -
0.017u, - 0.008t, which implies that wages in a long-run equilibrium (E(~~x,) = 0)
increase by 3 percent per annum after the influence of prices and unemployment
have been taken into account. Figure 4.10 gives the graphs of the disequilibria or
cointegrating vectors ~:x" the "actual" (Xii) and "fitted" (~fi'i f3jjXj,) values for x; =
150 MACROECONOMETIUCS
pJ.= _
2,---------, J..2,....---------,
.8
.4
:rrl:J=
6.------------, J..2 ,...~JI2::=~=== __-,
.8
9
J.975 J.985 J.995 J,989 J,999
vecto.. 3= _ _ 10=
J..2
.8
.4
.~
9
J.975 J.985 J.995 J.989 J.999
(w" PI' u,), and the recursively estimated eigenvalues (see Hansen and Johansen,
1993) each after having partialled out the full-sample short-run dynamics and the
unrestricted variables (that is, the constant and the dummies Budget, Expans;on,
D745, and D793). The first cointegrating vector appears to be /(0), but the other two
vectors are clearly nonstationary. There is a very close correspondence between
the "actual" and the "fitted" values for w,consistent with the disequilibrium rarely
being over 4 percent. Note, though, that the estimated coefficient of u, in the error
correction has the opposite sign to that estimated by Clements and Mizon (1991),
LSE METHODOLOGY lSI
implying that ceteris paribus wages increase with increases in unemployment and
so are counter cyclical! However, Clements and Mizon (1991) included average
labor productivity in their analysis and found evidence of a long-run positive
association between productivity and unemployment.
Note that these inferences concerning the integration-cointegration properties
of the system require the system parameters to be constant if they are to be valid,
whereas conventional test statistics for parameter constancy (such as system ana-
logues of the analysis of variance and Chow statistics mentioned in the previous
section) have known distributions for 1(0) rather than I( I) systems. Although con-
stancy tests for 1(1) variables are being developed, recursive estimation already
provides a valuable check on parameter constancy. It is therefore reassuring that
the recursive estimates of the eigenvalues are essentially constant. In particular,
the largest III is approximately 0.4 for sample sizes 35 to 109, with only a sugges-
tion of nonconstancy around 1979 and 1980. Of the other two eigenvalues, neither
of which is significantly different from zero in the full sample, J12 take values
declining from 0.2 to 0.1, and f.l3 is uniformly close to zero for all sample sizes.
Therefore, it was decided to proceed to analyze the system further on the assump-
tion that there is a single cointegrating vector.
Although the single cointegrating vector is already identified, it is of interest to
test overidentifying restrictions on it such as (1) the absence of an unemployment
effect on wages and prices ({313 = 0), (2) a long-run real wage equilibrium ({312 =
-I), and (3) no long-run trend in wages and prices ({314 = 0). In addition, necessary
conditions for P, and u, to be weakly exogenous for the parameters of the long-run
wage-price equation-(a21 = 0) and (a31 = 0), respectively-can be tested (see
Boswijk, 1992; Hendry and Mizon, 1993; Johansen, 1992a, 1992b; Urbain, 1992).
The test statistics are conventional likelihood ratio statistics, since the hypotheses
are linear on an 1(0) parameterization of the system (N.B. if X, - 1(1), ax, and {3'x,
- 1(0) when there is cointegration), and so they have limiting X 2 distributions with
the degrees of freedom equal to the number of independent restrictions being
tested.
The hypotheses that the long-run equilibrium is one for real wages ({312 = -1),
that prices are weakly exogenous for a long-run wage equation (a 21 = 0), that
wages are weakly exogenous for a long-run price equation (all =0), and that there
is no trend in the long-run wage-price equation (/314 = 0) are all rejected. However,
the hypotheses that u, is weakly exogeneous for the parameters of the long-run
wage-price equation (a 31 = 0), that there is no unemployment effect in the long-
run wage-price equation ({3J3 = 0), and that the adjustment coefficient for wages
is equal in magnitude but opposite in sign to that of prices (all + a 12 = 0) are not
rejected separately or jointly. Hence, the full-sample estimates of this system are
reasonably congruent and consistent with the existence of a long-run wage-price
equilibrium of the form
152 MACROECONOMETRICS
_o:k&;; I J~'~-~:~
.19'75 .1985 .1995 .19'75 .1985 .1995 .19'75 .1985 .1995
.~.
2
.,..r.. I"E'~
=:~~~~=~ =::
8
.197:1
=:.1985
A.2
.19'5
8'
Figure 4.11. /(0) system (PVAR) actual, fitted, and residual plots
dW" dp" and !:1u" and the associated residual correlograms and frequency plots in
Figure 4.11. As a result of imposing a rank of one on the cointegrating space
(which the statistics in Table 4.11 indicated to be appropriate), and dropping
the trend (other than its presence in the cointegrating vector ecm,), the /(0) system
has nine fewer parameters than the original system and so will be referred to as
a parsimonious VAR (PVAR). Although there is further scope for simplifying this
PVAR since a number of the estimated coefficients do not appear to be signifi-
cantly different from zero, the PVAR is retained in this form as a framework with-
in which to compare alternative models by checking their ability to parsimoniously
encompass the PVAR. The advantages of testing for simple models being able to
parsimoniously encompass the PVAR are (1) reducing the risk of using models
that are not robust to changes in the sample information (overparameterized models
can be too finely tuned to the peculiarities of a particular sample), (2) increasing
the chance that developed models are invariant to regime changes and thus cap-
turing autonomous relationships rather than ephemeral occurrences, and (3) allow-
ing the evaluation of models of particular interest (such as those implementing
specific economic theories).
154 MACROECONOMETIUCS
0.030793 - 1.10
(0.01) (0.48)
vecAR(45,253) 1.34 w, p, u,
vecH(156, 413) 1.04 p, -om 1.0
vecN(6) 7.45 u, O.ll 0.07 1.0
a a
_~:kk:;; I:~FE~~~L:~
BIoI Actual BIoI Res iduals Dp Actual
and ARCH effects in the residuals the validity of this parsimonious encompassing
test is called into question. However, comparison of Figure 4.12 with Figure 4.8
reveals the similarity between the congruent system and the simplified model and
so lends some support (albeit weak support) to the use of this likelihood ratio test
statistic.
The second model considered for encompassing comparison within the frame-
work of the PVAR is a VAR for the differences .1w" .1p" and .1u, (DVAR). This
model corresponds to (4.34) with a= 0 so that the disequilibrium ecm, is ignored.
As a system, a VAR in differences would be (4.31) with II = 0 so that the long-
run or zero-frequency information in the data is ignored and represents a class of
model that has been popular in the time series analysis of nonstationary data
particularly since the work of Box and Jenkins (1970). In addition to the three
restrictions in a = 0 there are a further six from restricting the number of dummy
variables included in each equation as in the simplified model. The likelihood ratio
test statistic takes the value ,t(9) = 45.06 [p = 0.00] so that the DVAR does not
parsimoniously encompass the PVAR. Note that relative to the test statistics re-
ported in Table 4.13 the present test is of the joint hypothesis ail = 0, i = 1, 2, 3,
plus the zero restrictions on the coefficients of the dummy variables. Hence, the
156 MACROECONOMETItlCS
fact that the hypothesis a 21 = 0 was rejected contributes to, but does not entirely
explain, the rejection of the DVAR model. The DVAR also fails to parsimoni-
ously encompass the PVAR when the dummy variables are unrestricted-x 2(3) =
33.86 [p = 0.00] and so within-sample the zero frequency information con-
tained in the cointegrating vector ecm, has a valuable role in the modelling w" PI'
and ",.
Recalling the evidence of a regime shift in 1979 for the single-equation modeling
in the previous section, it is relevant to test the parameter constancy of the other-
wise congruent system, the PVAR, the simplified model, and the DVAR. Table
4.16 presents the one-step ahead forecast test statistics from estimating the system
and models with data up to 1979(2) and checking their ability to forecast over the
fifty-five quarters in the period 1979(3)-1993(1). The three statistics are F j (l65,
T - k) = e''l'"(elI65, i = 1,2,3, when e = vec(E), E is a 55 x 3 matrix of forecast
errors for the three modeled variables, the 'l';'s are alternative estimates of the
asymptotic covariance matrix of e, T (= 53) the sample size, and k the average
number of estimated parameters per equation. 'l'1 is the sample estimate of the
innovation covariance matrix L, 'l'2 is an estimate of the forecast error covariance
matrix that allows for parameter uncertainty as well as the innovation variance,
and 'l'3 is an estimate of the forecast error covariance matrix that allows for inno-
vation variance, parameter uncertainty, and the covariance between the forecast
LSE METHODOLOGY 157
_·:t~1.:~ 1975
Dp Actual
1985
a Fitted
1995 1988
Dp Actual
1985 1998
a Forecast
1995
_·:[:4f!~;;1~~1
1975 1985 1995 1988 1985 1998 1995
-·:~I-·:~
-.4
1975 1985 1995
-.4
1988 1985 1998 1995
errors (see Doornik and Hendry, 1994 for details). The three statistics F;(l65, 53
=
- k), i I, 2, 3, have approximate central F distributions with degrees of freedom
165 and (53 - k) under the null hypothesis of parameter constancy.
The forecast perfonnance of the VAR is poor (it is only F)(165, 37) that does
not reject parameter constancy) as the graphs in Figure 4.13 make clear, especially
for I1w" which has the actual and forecast values diverging disconcertingly, and
the forecasts lying outside the confidence interval given by ±2 forecast error
standard deviations. The difficulties of the full-rank VAR are even more pro-
nounced when the forecasts are presented for the levels of the variables w,.
P" and
U,. However, the PVAR, the simplified model, and the DVAR all forecast much
better than the VAR, as evidenced by the statistics in Table 4.16. On the basis of
the parameter constancy test statistics and the means and standard deviations of
the forecast errors given in Table 4.16 the best overall perfonnance is found in the
simplified model, which is parsimonious and retains the zero frequency infonna-
tion in the disequilibrium ecm,. The improved quality of these forecasts relative to
those from the VAR is shown in Figure 4.14, which gives the actual, fitted, and
forecast values for the simplified model.
158 MACROECONOMETRICS
_·l~r:~ 1'75
Dp Actual a
1'85
Fitted
19'5 19811
Dp Actual
1985 1"11
a Forecast
1"5
:·~Ert·::3_.:~
Figure 4.14. Simplified model actual, fitted, and forecasts
Indeed, the difficulty lies in the different information used by the different
models in forecasting the variables W" P" and U ,• The VAR uses the sample infor-
mation on the levels of W" PI' and U , and their history, but wrongly treats the VAR
as having full rank, and also ignores the regime shift in 1979. The DVAR uses
sample information on the differences dw dp" and du, and their history but
ignores sample information on the levels of "these variables, the fact that the VAR
has reduced rank, and that there is a regime shift. The PVAR (that is, the reduced-
rank VAR mapped into /(0) space) and the simplified model, on the other hand,
implicitly use sample information on the levels of the variables and their history,
as well as the fact that the VAR does not have full rank, but ignore the regime shift.
However, the latter two models forecast well because they use a full sample
estimate of ecm which thus reflects the regime shift and so keeps the forecasts on
track. If instead" ecm, were replaced by the estimate of the cointegrating vector
using data from 1966(1) to 1979(2) only, the forecast performance of all the
models (other than the VAR in differences) deteriorates to be similar to that of
the VAR. Table 4.17 gives the split sample unrestricted (ecm,) estimates of the
cointegrating vector, plus restricted estimates (ecm~) satisfying the nonrejected
overidentifying restrictions discussed below.
LSE METHODOLOGY 159
1979(3)-1993(1)
* Significant at
** Significant at
in Table 4.18, which is very different from the post-1979(2) error correction ecm2~
estimated subject to the restrictions a 31 = 0, a 21 = 0, 1313 = 0, and 1312 = -1 (X 2(4)
= 8.40 [p > 0.08]). Hence there is evidence that U, plays no role in the determi-
nation of the long-run equilibrium values of W, and P, after 1979(2) and that W,
adjusts fully to P, and in addition has an autonomous trend. Finally, notice that
ecm, and ecm2, (ecmf and ecm2~) are very similar and noticeably different from
ecml , (ecml~). This helps to understand why the one-step-ahead forecasts for the
period 1979(3)-1993(1) from the PVAR and the simplified model, both of which
have ecm, as an explanatory variable, are better than the corresponding forecasts
from the VAR. In fact, if the single-equation model for wages is reparameterized
to be in /(0) space, it too has good one-step-ahead forecasts. In particular, if
WH , P,-I, U,_I' and t in the reduced model reported in Table 4.3, or W,_I, P,-I' and
t in the final model reported in Table 4.4, are replaced by ecm ,_1 the resulting
forecasts for ~W, are very similar to those for ~W, of the simplified model shown
in Figure 4.14. However, the fact that P, is not weakly exogenous for the para-
meters of the long run equilibrium E(ecm ,) = 0 means that more efficient infer-
ences should result from jointly modelling W, and P" rather than using single
equation OLS analysis. In fact, the forecasts for W, from the system modeling
should be preferable to those from the single-equation modeling. This is borne out
in practice as can be seen from comparison of the forecast error mean and standard
deviation from the single equation analysis (-0.0167 and 0.0168, respectively)
with those for the simplified model reported in Table 4.16.
LSE METHODOLOGY 161
Although the above empirical analysis illustrates the importance of testing for
weak exogeneity and modeling the joint distribution of variables in its absence, it
is important to note that reasonable forecasts for the period 1966( 1)-1993(1) were
only obtained by using the full-sample estimate of the disequilibrium or error cor-
rection ecm" which would not be possible in practice for ex ante forecasting. The
only model considered in this paper that would have been successful in ex ante
forecasting is the DVAR. This highlights the fact that the presence of regime shifts
in integrated-cointegrated systems for macroeconomic time series presents a sub-
stantial challenge to econometric modeling. Differencing time series that are
subject to regime shifts does not account for the shifts even when they might be
identifiable as changes in structure resulting from changes in government eco-
nomic policies, but joint modeling of the differences in a VAR does enable the
generation of multistep forecasts that converge on the unconditional means of the
series (see Hendry and Clements, 1994). Forecasts from econometric models that
include variables that have not been suitably differenced to transform regime
shifts from step or trend changes into impulses or blips that have no duration (such
as ecm, in the above illustration) may require intercept corrections or other adjust-
ments to keep them on track when there are regime shifts. Seen in this light the
use of intercept corrections can be interpreted as a means of exploiting past fore-
cast errors to keep present forecasts on track (see Clements and Hendry, 1994).
It is instructive also to note that the long run disequilibrium variable ecm"
despite the fact that it is commonly called an error-correction mechanism, does not
error-correct when it does not incorporate an explanation for regime shifts that are
present in the system being modeled. Hence, ecm, and E(ecm,) = 0 have been
referred to as long-run disequilibrium and long-run equilibrium, respectively.
Ideally, econometric models should provide explanations for structural changes
and account for autonomous shifts that affect the variables being modeled and thus
avoid the necessity of differencing such changes out of the system in order to
produce reliable forecasts. Although this statement describes an enormous chal-
lenge to econometric modelers, those who overcome the challenge are likely to
develop econometric models with great potential for use in economic policy analysis
and forecasting. To rise to this challenge in the modeling of wages, prices, unem-
ployment, and other related variables such as productivity over the period 1966(1)-
1993(1) in the United Kingdom, it is likely that account will have to be taken of
the important changes in the labor market structure, in exchange-rate regimes, and
government policies toward inflation and unemployment. The fact that the models
presented in this section have a single deterministic trend to approximate the
effects of technical change, changes in productivity, and external factors and
do not have long-run equilibria with price homogeneity (except for ecm2~ in
Table 4.17) indicates that there is ample scope for further modeling. This impor-
tant task lies beyond the scope of the present paper, which has concentrated on the
162 MACROECONOMETRICS
Conclusions
The LSE modeling methodology developed rapidly during the 1960s and 1970s
and has had a significant effect on econometric modeling at large, as described
in Gilbert (1986). However, no vital methodology is static but rather evolves
with important developments. This is certainly true of the LSE methodology, and
as suggested by Pagan (1992) a more accurate name today might be the LSE-
Copenhagen-San Diego-Yale methodology-though even this ignores the valu-
able contributions of researchers in institutions in other geographical locations.
The econometric modeling methodology associated with the LSE has achieved
much, is still evolving and being refined, and provides a powerful tool for reaping
the benefits of scientific econometric modeling. Though ultimately the model
is the message, the system (that is, the congruent representation of the joint dis-
tribution of the observed variables of relevance) is the appropriate statistical frame-
work for developing and evaluating such econometric models. The legacy of this
methodology includes the widely accepted need to rigorously evaluate econo-
metric models by checking their congruence with available information and their
ability to encompass rival models, as well as the advocacy of the general-to-
specific modeling strategy as an efficient and efficacious modeling strategy.
Econometric concepts that provide important objectives to be achieved in the
LSE methodology are homoscedastic innovation error processes; weak, strong,
and super exogeneity; encompassing; and parameter constancy. Failure to achieve
these objectives in practice can result in econometric models that are likely to
yield misleading conclusions from economic policy analyses and produce unreli-
able forecasts. Indeed one of the most important challenges in the econometric
modeling of macroeconomic time series is to produce multistep forecasts that are
at least as accurate as those generated from multivariate time-series models such
as DVARs. This is a goal to which many of those espoused to the LSE method-
ology aspire. The simple illustration of empirical modeling presented above dem-
onstrates the seriousness of this challenge but also suggests that there is hope of
future success. In addition, the illustration highlighted the important effect changes
in economic policy can have on the inter-relationships between macroeconomic
variables, especially on the underlying equilibrating mechanisms and the exogeneity
status of variables within them. The changes in economic policy from monitoring
the rate of unemployment, to reducing radically the rate of inflation, had major
effects on the U.K. economy post-1979.
LSE METHODOLOGY 163
Data Appendix
The quarterly data set runs from 1965(1) to 1993(1). The precise estimation pe-
riods used vary according to the number of lags in the specification. All series are
seasonally adjusted except for the price series. The wage variable w is defined as
log (WS/EE* A VH), where WS is wages and salaries, EE is employees in employ-
ment, and A VH is a measure of average weekly hours in the manufacturing sector.
The price variable p is the log(P) when P is the retail price index, so that t:.p is
the quarterly aggregate inflation rate. Finally, u =10g(U), U being the percentage
of unemployment rate.
The precise definitions and sources are as follows:
WS:
Wages, salaries and forces pay. imn. ETAS, 65:1 to 90:4
CSO mnemonic ADB ET, 91:1 to 93:1
EE:
Employees in employment OOO's, whole UK ETAS, 65:1 to 89:3
CSO mnemonic BCAl ET, 89:4 to 93: 1
AVH:
Index of average weekly hours worked per EG, 65: 1 to 93: 1
operative in manufacturing (1980 = 100)
P:
Retail price index, all items (1987 = 100) ETAS, 65:1 to 92:2
CSO mnemonics FRAG and CHAW ET, 92:3 to 93: 1
U:
Unemployment rate (UK) percent ETAS, 65:1 to 91:1
CSO mnemonic BCJE ET, 91:2 to 93:1
Acknowledgments
The views presented in this paper on the nature and role of econometric modeling
are mine. However, since I am associated with what has become known as the
LSE methodology (or sometimes Professor Hendry's methodology) (see Gilbert,
1986; Ericsson, Campos, and Tran, 1990), this is consistent with my brief to write
164 MACROECONOMETRICS
a paper describing the LSE methodology. I should like to thank my many friends
and colleagues of the "LSE School" for the intellectual stimulation they have
provided over the last quarter of a century. In particular, I wish to thank Mike
Clements, David Hendry, Maozu Lu, and Jean-Fran~ois Richard for allowing me
to draw on our joint research without wishing to hold them responsible for the
result. I am also grateful to David Hendry for providing many valuable sugges-
tions while the paper was in gestation and to Mike Clements and Kevin Hoover
for providing detailed and constructive comments on an earlier version of the
paper. Financial support for this research from the ESRC under grant ROOO231184
and from the EUI Research Council is gratefully acknowledged.
References
Kydland, F., and E. Prescott. (1982). "Time to Build and Aggregate Fluctuations."
Econometrica 50, 1345-1370.
Kydland, F., and E. Prescott. (1995). "The Econometrics of the General Equilibrium Ap-
proach to Business Cycles." In K.D. Hoover (ed.), Macroeconomics: Developments,
Tensions and Prospects (ch. 5). Boston: Kluwer.
Layard, R., and S.J. Nickell. (1985). "The Causes of British Unemployment." National
Institute Economic Review 111,62-85.
Leamer, E.E. (1978). Specification Searches: Ad Hoc Inference with Non-Experimental
Data. New York: Wiley.
Leamer, E.E. (1983). "Let's Take the Con Out of Econometrics." American Economic
Review 73, 31-44.
Lipsey, R.G. (1960). "The Relation Between Unemployment and the Rate of Change of
Money Wage Rates in the United Kingdom 1862-1957: A Further Analysis." Economica
27, 1-31.
Lu, M., and G.E. Mizon. (1991). "Forecast Encompassing and Model Evaluation." In P.
Hackl and A. Westlund (eds.), Economic Structural Change: Analysis and Forecasting
(pp. 123-138). Vienna: Springer-Verlag.
Lucas, R.E. (1976). "Econometric policy evaluation: a critique." In K. Brunner and A.H.
Meltzer (eds.), The Phillips Curve and Labor Markets (Vo!. I) (pp. 19-46). Carnegie-
Rochester Conference Series on Public Policy. Amsterdam: North Holland.
Maasoumi, E. (1988). Contributions to Econometrics: John Denis Sargan (Vols. 1-2).
Cambridge: Cambridge University Press.
Mizon, G.E. (1977a). "Model Selection Procedures in Dynamic Models." In M.J. Artis and
A.R. Nobay (eds.), Studies in Modern Economic Analysis (ch. 4, pp. 97-120). Oxford:
Basil Blackwell.
Mizon, G.E. (1977b). "Inferential Procedures in Nonlinear Models: An Application in a UK
Cross Sectional Study of Factor Substitution and Returns to Scale." Econometrica 45,
1221-1242.
Mizon, G.E. (1984). "The Encompassing Approach in Econometrics." In D.F. Hendry and
K.F. Wallis (eds.), Econometrics and Quantitative Economics (ch. 6, pp. 135-172).
Oxford: Basil Blackwell.
Mizon, G.E. (1989). "The Role of Econometric Modelling in Economic Analysis." Revista
Espanola de Economia 6, 165-191.
Mizon, G.E. (1991). "The Role of Measurement and Testing in Economics." In M. Bleaney,
D. Greenaway, and I.M. Stewart (eds.), Companion to Contemporary Economic Thought
(ch. 28, pp. 547-592). London: Routledge.
Mizon, G.E. (1995). "A Simple Message for Autocorrelation Correctors: DON'T." Journal
of Econometrics, forthcoming.
Mizon, G.E., and D.F. Hendry. (1980). "An Empirical Application and Monte Carlo Analy-
sis of Tests of Dynamic Specification." Review of Economic Studies 49, 21-45.
Mizon, G.E., and J-F. Richard. (1986). "The Encompassing Principle and Its Application
to Non-Nested Hypothesis Tests." Econometrica 54, 657-678.
Monfort, A., and R. Rabemananjara. (1990). "From a VAR Model to a Structural Model:
With an Application to the Wage-Price Spira!." Journal of Applied Econometrics 5,
203-227.
LSE METHODOLOGY 169
Nickell, SJ. (1984). "The Modelling of Wages and Employment." In D.F. Hendry and K.F.
Wallis (eds.), Econometrics and Quantitative Economics (ch. 2, pp. 13-35). Oxford:
Basil Blackwell.
Osterwald-Lenum, M. (1992). "A Note with Quantiles of the Asymptotic Distribution of
the ML Cointegration Rank Test Statistics." Oxford Bulletin ofEconomics and Statistics
54,461-472.
Pagan, A.R. (1987). "Three Econometric Methodologies: A Critical Appraisal." Journal of
Economic Surveys 1, 3-24. Reprinted in C.WJ. Granger (ed.), Modelling Economic
Series: Readings in Econometric Methodology (pp. 97-120). Oxford: Oxford Univer-
sity Press.
Pagan, A.R. (1989). "Twenty Years After: Econometrics, 1966-1986." In B. Comet and
H. Tulkens (eds.), Contributions to Operations Research and Econometrics: The Twen-
tieth Anniversary of CORE (pp. 319-383). Cambridge, MA: MIT Press.
Pagan, A.R. (1992). "Three Econometric Methodologies: An Update." Mimeo, Australian
National University.
Phillips, A.W.H. (1958). "The Relation Between Unemployment and the Rate of Change
of Money Wages Rates in the United Kingdom 1861-1957." Economica 25, 283-299.
Phillips, P.C.a. (1988). "The ETlnterview: A.R. Bergstrom." Econometric Theory 4, 301-
327.
Phillips, P.C.B. (ed.). (1993a). Models. Methods. and Applications ofEconometrics: Essays
in Honor of A.R. Bergstrom. Cambridge: Basil Blackwell.
Phillips. P.C.B. (l993b). "Rex Bergstrom's Career and Research." In P.C.a. Phillips (ed.),
Models. Methods, and Applications ofEconometrics: Essays in Honor ofA.R. Bergstrom
(pp. 3-8). Cambridge: Basil Blackwell.
Pesaran, M.P. (1974). "On the General Problem of Model Selection." Review ofEconomic
Studies 41, 153-171.
Prothero, D.L., and K.F. Wallis. (1976). "Modelling Macro-Economic Time Series (with
discussion)." Journal of the Royal Statistical Society Series A, 139,468-500.
Reimers, H.-E. (1992). "Comparisons of Tests for Multivariate Cointegration." Statistical
Papers 33, 335-359.
Richard, J-F. (1980). "Models with Several Regimes and Changes in Exogeneity." Review
of Economic Studies 47, 1-20.
Sargan, J.D. (1964). "Wages and Prices in the United Kingdom: A Study in Econometric
Methodology." In P.E. Hart, G. Mills, and J.K. Whitaker (eds.), Econometric Analysis
for National Economic Planning. London: Butterworths. Reprinted in D.F. Hendry and
K.F. Wallis (eds.), Econometrics and Quantitative Economics (pp. 275-314). Oxford:
Basil Blackwell.
Sargan, J.D. (1980a). "A Model of Wage-Price Inflation." Review ofEconomic Studies 47,
97-112.
Sargan, J.D. (1980b). "Some Tests of Dynamic Specification for a Single Equation."
Econometrica 48, 879-897.
Sargan, J.D. (1988). Lectures on Advanced Econometric Theory. Oxford: Basil Blackwell.
Sims, c.A. (1980). "Macroeconomics and Reality." Econometrica 48, 1-48. Reprinted in
C.WJ. Granger (ed.), Modelling Economic Series: Readings in Econometric Method-
ology (pp. 137-189). Oxford: Oxford University Press.
170 MACROECONOMETIUCS
Sims, c.A. (1992). "Interpreting the Macroeconomic Time Series Facts: The Effects of
Monetary Policy." European Economic Review 36, 975-1000.
Spanos, A. (1986). Statistical Foundations of Econometric Modelling. Cambridge: Cam-
bridge University Press.
Spanos, A. (1988). "Error Autocorrelation Revisted: The AR(I) Case." Econometric Re-
views 6, 285-294.
Summers, L.H. (1991). "The Scientific Illusion in Empirical Macroeconomics." Scan-
dinavian Journal of Economics 93, 129-148.
Terasvirta, T. (1970). "Stepwise Regression and Economic Forecasting." Economic Stud-
ies Monograph No. 31. Helsinki: Finnish Economics Association.
Urbain, J.P. (1992). "On Weak Exogeneity in Error Correction Models." Oxford Bulletin
of Economics and Statistics 54, 187-208.
Vernon, K. (1970). "An Econometric Study of Wage and Price Inflation in the United
Kingdom in the Post War Period." Ph.D. thesis, London School of Economics.
West, K.D. (1988). "Asymptotic Normality When Regressors Have a Unit Root." Econo-
metrica 56, 1397-1417.
White, H. (1980). "A Heteroskedastic-consistent Covariance Matrix Estimator and a Direct
Test for Heteroskedasticity." Econometrica 48, 817- 838.
Zellner, A., and F. Palm. (1974). "Time Series Analysis and Simultaneous Equation Econo-
metric Models." Journal of Econometrics 2, 17-54.
Commentary on Chapter 4
Jon Faust
Charles H. Whiteman
Introduction
Over the past two decades or so, many approaches to the modeling of macro-
economic time series have come and gone; one that has come and stayed is the
LSE approach. Advocates of the approach have been prolific in asserting the
virtues of their econometrics and the shortcomings of what Gilbert (1986) calls
the average economic regression (AER) (Doomik and Hendry, 1994; Ericsson,
Campos, and Tran, 1990; Gilbert, 1986, 1989; Hendry, 1987, 1995; Spanos, 1986).
Despite these writings, the average econometrician may sympathize with a hotly-
pursued Robert Redford (as the Sundance Kid), who repeatedly asked, "Who are
those guys?"
Grayham Mizon approaches this question from several directions: he provides
a general description of the methodology; he lists numerous contributors to the
development of the methodology; he describes its main components; and he pro-
vides an example.
Having addressed the question, has Mizon answered it? We are not sure. Much
of Mizon' s discussion is couched in the special vocabulary of the LSE approach,
making it difficult to determine what (other than language) distinguishes those
guys from, say, the average econometrician reading this volume. Thus, we set out
to distill from Mizon's paper what is unique about the LSE methodology.
The place to begin trying to understand the LSE approach is at the beginning,
Mizon's introduction. Therein, he states, "The essence of this [LSE] approach is
the recognition that potentially valuable information for the analysis of any eco-
nomic problem can come from numerous sources including, economic theory, the
available sample of observations on the potentially relevant variables, knowledge
of the economic history of the period under study, and from knowledge of the way
in which the observed data are defined and measured, plus their relationship to the
theory variables."
If this is true, then the LSE approach must share its essence with that of the
171
172 MACROECONOMETRlCS
The LSE has been a center of tremendous productivity in econometrics, and the
list of econometricians who have exited the tube at Holborn as faculty or visitors
at the nearby LSE is a Who's Who of econometrics. Can one understand what is
called the LSE approach by looking at the work of those individuals identified as
having had important influence on it? No. There may be common features in the
work of as diverse a group of eminent econometricians as James Durbin, Peter
Phillips, Kenneth Wallis, Jean-Francois Richard, and Takeshi Amerniya, but those
common features do not add up to the LSE approach. Many of those on Mizon' s
list do not even use the LSE methodology in their research. Of course, the LSE
approach has benefitted from a large and diverse group of econometricians; per-
haps it is more than the sum of its parts. To the parts we turn.
The goal of LSE econometrics, as described and applied in the third and fourth
sections, seems to be to fit a congruent, encompassing model to the DGP;
COMMENTARY ON CHAPTER 4 173
DGP
The DGP is what other econometricians might refer to as the "likelihood function"
or the ')oint distribution." Mizon notes that there are many factorizations and
observationally equivalent parameterizations of the likelihood. Though all
factorizations are equal, some are more equal than others. To wit (third section)
"Now consider another investigator who claims to be interested in the model:
[M:J ... this investigator is clearly not considering the 'true' model since M 2 does
not correspond to any of the equations of the DGP as given in (2)! It is, though,
part of other valid parameterizations of the DGP ... and thus 'true.'"
It is clear that the nature of the LSE approach is tied to its users' interpretations
of the meaning of words like true and valid, meanings that to us are not altogether
clear. Meanwhile, our (not-so-clear) understanding of these statements and the
ones surrounding them is that different factorizations and parameterizations of the
likelihood may be of focal interest depending on the perspective and purpose. If
we are correct, this point is not unique to the LSE approach.
Mizon's discussion does reveal one feature that distinguishes the LSE approach
from some others: it is likelihood-based and pays great attention to what different
factorizations of the likelihood imply for the properties of inferences. In this re-
spect, the approach is founded in the statistical literature on the related concepts
of sufficient and ancillary statistics, and admissibility and complete classes of
procedures (see Barndorff-Nielsen, 1978; Basu, 1977;, Lehmann, 1986). Engle,
Hendry, and Richard (1983), for example, applied the notion of ancillarity in
defining weak exogeneity.
Congruence
Being congruent means being consistent, in some sense, with infonnation from
four sources: a priori theory, sample infonnation, the measurement system, and
rival models.
There is a spectrum of views in the profession on how big a role theory should
play. Everything from nothing (VARs) to everything (RBC modelers are "ahead
174 MACROECONOMETIUCS
of measurement," Prescott, 1986). Where does the LSE school fall? Mizon cites
Hendry as saying that under the LSE approach, consistency with economic theory
is a desirable but inessential property. Apparently, measurement without (much)
theory is permissible. Indeed, we do not know many economic theorists who
would find measurable comfort in Mizon's Table 4.14. Theory consistency is not
a particularly important component of congruence.
In principle, the model one uses should impose such things as accounting identi-
ties and nonnegativity constraints. Pragmatic application (or not) of this restriction
as advocated by Mizon seems common to most approaches.
Parsimonious Encompassing
The LSE has been at the forefront of advocating that new econometric models
dominate those that preceded and has been instrumental in developing encompass-
ing tests for evaluating the relative merits of econometric models (see Mizon,
1984). Comparing new models to old ones is surely part of any progressive ap-
proach, and the formal merits of encompassing tests seem well accepted.
Suppose one had a model that passed all the encompassing tests one had con-
ducted. What would one have? Mizon claims that "an encompassing model is one
that can account for the previous empirical findings that were thought to be rel-
evant and adequate for the explanation of the variables and parameters of interest,
and can explain the features of rival models being considered currently." Rel-
evance to whom? adequate in what way? explain how? which features? Mizon
elaborates that the encompassing model will be the "dominant model that is con-
gruent with information from allfour sources" (emphasis added). Since one of the
four sources is a priori theory, which is inessential, this statement should probably
read from "three of four sources." In what sense is the model dominant? Mizon
explains that "the encompassing model renders other models inferentially redun-
dant and so is a dominant model." Based on the discussion of validity above, we
interpret this as a claim that restricting attention to the encompassing model
involves no loss of infonnation relative to considering a larger set of models or
information. The literature on admissible procedures cited above provides a natu-
ral formal interpretation of this claim: an admissible decision procedure is, loosely
speaking, one that is best under some loss function. One might say model A
renders model B inferentially redundant if the set of admissible decision proce-
dures based on model A contains all of the admissible procedures based on models
A and B.
But parsimonious encompassing does not deliver a model that is dominant in
this sense. For a simply exposited but trivial example, consider using standard
tests as advocated by Mizon to select variables for two parsimonious models of
the same data. For model A use I percent critical values; for model B use 10 per-
cent critical values. Generally, model B will be nested in model A: it is harder to
reject exclusion restrictions at the I percent level. Since it is nested and justified
by 10 percent tests, model B will parsimoniously encompass the model A (at the
10 percent level). Decisions based on information in model A but not in model B
will generally be admissible, however. Similar examples can be constructed for
nonnested tests and tests of maintained restrictions. So sometimes one might want
to use an encompassed model.
Very generally, parsimonious encompassing and dominance are like validity-
Good Things. But it is the specifics about which we are unsure: a translation of
LSE terms like validity, no loss of infonnation, dominance, and inferential
176 MACROECONOMETRICS
General-ro-Specific Modeling
Of all the elements of the approach, this final element of the LSE approach is the
most clear. Start with a general model and obtain parsimony by applying and
retaining restrictions that are not rejected in classical tests.
Having set out to distill the essence of the LSE approach, we think it is fair to
say that our understanding of its distinguishing features has attained a clarity
falling somewhere between that of Newcastle Brown and Old Milwaukee. We
now tum to our opinion of this brew.
We have tried to specify some gross principles of the LSE approach that collec-
tively set it apart from other approaches to modeling macroeconomic time series.
Our brief list indubitably glosses over some important contributions of the ap-
proach, most notably, perhaps, the work cited by Mizon on dynamic specification,
error correction, and cointegration.
The most general underlying principle seems to be, "If something is testable,
test it," or simply "Test, test, and test" (Hendry, 1980). We see five guidelines
filling out this dictum:
senses is unclear. Second, statistics, in our view, should be concerned with how
best to use information, not with mandating its full use. Since using information
can be difficult or costly, full use of information need not be (economically)
efficient. In short, we question the goal of no loss of infonnation and wonder how
the LSE approach attains it.
We also find understanding the role of economic theory in the LSE approach
problematic (the fifth guideline). As Mizon notes, Hendry has called consistency
with theory inessential. Doornik and Hendry (1994) call theory essential but reach
the conclusion that "overall, one can do little better than state the need for an
econometric model to be theory consistent." In what sense?
Despite the fact that LSE econometricians have done considerable work on
applying long-run homogeneity restrictions from theory using error correction
models and cointegration (see Hendry, 1995; BanneIjee et al., 1993), it seems to
us that evidence has the upper hand in the LSE approach. The general model with
which one starts is supposed to nest, at least in some approximate sense, theories
of interest, but the final model will retain a recognizable association with theory
only if the theory is favored by the evidence as divined through the prescribed
sequence of tests. This balance of theory and evidence makes perfect sense if one
believes that there is no a priori or inductive reason to believe in theory. While
this position is coherent and, perhaps, reasonable, very different ways of balancing
theory and evidence are also coherent and reasonable.
Conclusion
The LSE school has made profound contributions to the progressive modeling
of economic time series. In particular, the school stands out in its emphasis on
general-to-specific modeling, its emphasis on testing maintained assumptions, and
in its emphasis on new models dominating those that preceded. LSE
econometricians have made a strong mark on econometric practice by advocating
these principles and by developing tools and approaches for following them.
Given the difficulties faced in the modeling of economic time series-non-
experimental data and short correlated samples-it is difficult to make a case that
any single approach will foster the most progress in understanding the economy.
The balance between a priori theory and data is at the heart of the differences
among many of the popular approaches. It is clear to us that understanding of the
economy may be enhanced by work involving the LSE approach and other ap-
proaches placing even less emphasis on constraints from theory, as well as work
involving greater emphasis on theory. That there there are so many approaches
being pursued and so little agreement about which is best is evidence that progress
to date has not strongly favored a single progressive approach.
COMMENTARYONC~R4 179
References
K.F. Wallis (eds.), Econometrics and Quantitative Economics (pp. 135-172). Oxford:
Basil Blackwell.
Oliner, S., G. Rudebusch, and D. Sichel. (1995). "The Lucas Critique Revisited: Assessing
the Stability of Empirical Euler Equations for Investment." JOUT7Ul1 of Econometrics,
forthcoming.
Prescott, E.C. (1986). "Theory Ahead of Business Cycle Measurement." Federal Reserve
Bank of Minneapolis Quarterly Review IO (Fall), 9-22. Also in K. Brunner and A.
Meltzer (eds.), Real Business Cycles, Real Exchange Rates and Actual Policies (pp. 11-
45). Carnegie-Rochester Conference Series on Public Policy 25 (Autumn). Amsterdam:
North-Holland.
Sargan, 1.D. (1981). "The Choice Between Sets of Regressors." Manuscript. London School
of Economics.
Spanos, A. (1986). Statistical Foundations of Econometric Modelling. Cambridge: Cam-
bridge University Press.
5
THE ECONOMETRICS OF THE
GENERAL EQUILIBRIUM APPROACH
TO BUSINESS CYCLES*
Finn E. Kydland and Edward C. Prescott
I. Introduction
Early in this century American institutionists and members of the German histori-
cal school attacked-and rightfully so-neoclassical economic theory for not
being quantitative. This deficiency bothered Ragnar Frisch and motivated him,
along with Irving Fisher, Joseph Schumpeter, and others, to organize the Econo-
metric Society in 1930. The aim of the society was to foster the development of
quantitative economic theory-that is, the development of what Frisch labeled
econometrics. Soon after its inception, the society started the journal Econometrica.
Frisch was the journal's first editor and served in this capacity for twenty-five
years.
In his editorial statement introducing the first issue of Econometrica (1933),
Frisch makes it clear that his motivation for starting the Econometric Society was
the "unification of theoretical and factual studies in economics" (p. I). This uni-
fication of statistics, economic theory, and mathematics, he argues, is what is
powerful. Frisch points to the bewildering mass of statistical data becoming avail-
able at that time, and asserts that in order not to get lost "we need the guidance
and help of a powerful theoretical framework. Without this no significant interpre-
tation and coordination of our observations will be possible" (ibid., p. 2).
181
182 MACROECONOMETRICS
Frisch speaks eloquently about the interaction between theory and observation
when he says "theory, in fonnulating its abstract quantitative notions, must be in-
spired to a larger extent by the technique of observation. And fresh statistical and
other factual studies must be the healthy element of disturbance that constantly
threatens and disquiets the theorist and prevents him from coming to rest on some
inherited, obsolete set of assumptions" (ibid.). Frisch goes on to say that
this mutual penetration of quantitative economic theory and statistical observation is the
essence of econometrics (ibid., p. 2).
Tjalling Koopmans, who was influenced by Frisch and might even be considered
one of his students, argued forcefully in the late 1940s for what he called the
econometric approach to business cycle fluctuations. At the time, it was the only
econometric approach. The general equilibrium approach to the study of business
cycles had yet to be developed. But since the approach Koopmans advocated is no
longer the only one, another name is needed for it. As it is the equations which are
invariant and measured, we label this approach the system-of-equations approach.'
In the 1930s, there were a number of business cycle models or theories. These
logically complete theories were a dynamic set of difference equations that could
be used to generate time series of the aggregate variables of interest. Notable
examples include Frisch's (1933) model in Cassel's volume, Tinbergen's (1935)
suggestions on quantitative business cycles, and Samuelson's (1939) multiplier-
accelerator model. One problem with this class of models is that the quantitative
behavior of the model depended upon the values of the coefficients of the vari-
ables included in the equations. As Haberler (1949) points out in his comment on
Koopmans' (1949) paper, the stock of cyclical models (theories) is embarrassingly
large. Give any sophomore "a couple of lags and initial conditions and he will
construct systems which display regular, damped or explosive oscillation ... as
desired" (p. 85). Pure theory was not providing sufficient discipline, and so it is
GENERAL EQUILIBRIUM APPROACH TO BUSINESS CYCLES 185
not surprising that Koopmans advocated the use of the statistics discipline to
develop a theory of business fluctuations.
System-at-Equations Models
As Koopmans (1949, p. 64) points out, the main features of the system-of-
equations models are the following: First, they serve as theoretical exercises and
experiments. Second, the variables involved are broad aggregates, such as total
consumption, the capital stock, the price level, etc. Third, the models are "logi-
cally complete, i.e., they consist of a number of equations equal to the number of
variables whose course over time is to be explained." Fourth, the models are
dynamic, with equations determining the current values of variables depending
not only on current values of other variables but also on the values of beginning-
of-period capital stocks and on lagged variables. Fifth, the models contain, at
most, four kinds of equations, which Koopmans calls structural equations. The
first type of equations are identities. They are valid by virtue of the definition of
the variables involved. The second type of equations are institutional rules, such
as tax schedules. The third type are binding technology constraints, that is, pro-
duction functions. The final type are what Koopmans calls behavioral equations,
which represent the response of groups of individuals or firms to a common
economic environment. Examples are a consumption function, an investment
equation, a wage equation, a money demand function, etc. A model within this
framework is a system-of-equations. Another requirement, in addition to the one
that the number of variables equal the number of equations, is that the system
have a unique solution. A final requirement is that all the identities implied by the
accounting system for the variables in the model hold for the solution to the
equation system; that is, the solution must imply a consistent set of national
income and product accounts.
equations and the parameters of the disturbance distribution. The crucial point is
that the equations of the macroeconometric model are the organizing principle of
the system-of-equations approach. What is measured is the value of the coeffi-
cients of the equations. The criterion guiding the selection of the values of the
coefficients is essentially the ability of the resulting system of equations to mimic
the historical time series. The issue of which set of equations to estimate is settled
in a similar fashion. The criterion guiding the selection of equations is in large part
how well a particular set can mimic the historical data. Indeed, in the 1960s a stu-
dent of business cycle fluctuations was successful if his particular behavioral equa-
tion improved the fit of, and therefore replaced, a currently established equation.
With the emergence of a consensus on the structure of the system of equations that
best described the behavior of the aggregate economy, the approach advocated by
Koopmans became totally dominant in the 1960s. This is well-illustrated by the
following statement of Solow's, quoted by Brunner (1989, p. 197):
I think that most economists feel that the short run macroeconomic theory is pretty well
in hand.... The basic outlines of the dominant theory have not changed in years. All
that is left is the trivial job of filling in the empty boxes [the parameters to be estimated]
and that will not take more than 50 years of concentrated effort at a maximum.
A powerful theoretical framework was developed in the 1950s and 1960s that built
upon advances in general equilibrium theory, statistical decision theory, capital
GENERAL EQUILffiRIUM APPROACH TO BUSINESS CYCLES 187
One of these important questions, which has occupied business cycle theorists
since the time of Frisch and Slutzky, is how to detennine which sources of shocks
give rise to cycles of the magnitudes we observe. To provide reliable answers to
this and similar questions, abstractions are needed that describe the ability and
willingness of agents to substitute commodities, both intertemporally and intra-
temporally, and within which one can bring to bear statistical or factual infonna-
tion. One of these abstractions is the neoclassical growth model. This model has
proven useful in accounting for secular facts. To understand business cycles, we
rely on the same ability and willingness of agents to substitute commodities as
those used to explain the growth facts. We are now better able than Frisch was
more than 50 years ago to calibrate the parameters of aggregate production tech-
nology. The wealth of studies on the growth model have shown us the way. To
account for growth facts, it may be legitimate to abstract from the time allocation
between market and nonmarket activities. To account for business cycle facts,
however, the time allocation is crucial. Thus, good measures of the parameters of
household technology are needed if applied business cycle theory is to provide
reliable answers.
Question
To begin with, the research question must be clearly defined. For example, in
some of our own research we have looked at quantifying the contribution of
changes in a technology parameter, also called Solow residuals, as a source of U.S.
postwar business cycles. But we refined it further. The precise question asked is
how much variation in aggregate economic activity would have remained if tech-
nology shocks were the only source of variation. We emphasize that an econometric,
that is, quantitative theoretic analysis, can be judged only relative to its ability to
address a clear-cut question. This is a common shortcoming of economic modeling.
When the question is not made sufficiently clear, the model economy is often
criticized for being ill-suited to answer a question it was never designed to answer.
Model Economy
To address a specific question one typically needs a suitable model economy for
addressing the specified question. In addition to having a clear bearing on the
question, tractability and computability are essential in determining whether the
model is suitable. Model-economy selection depends on the question being asked.
Model-economy selection should not depend on the answer provided. Searching
within some parametric class of economies for the one that best fits some set of
aggregate time series makes little sense. Unlike the system-of-equations approach,
no attempt is made to determine the true model. All model economies are abstrac-
tions and are by definition false.
190 MACROECONOMETRICS
Calibration
Computational Experiments
Once the model is calibrated, the next step is to carry out a set of computational
experiments. If all the parameters can be calibrated with a great deal of accuracy,
then only a few experiments are needed. In practice, however, a number of experi-
ments are typically required in order to provide a sense of the degree of confidence
in the answer to the question. It often happens that the answer to the research ques-
tion is robust to sizable variations in some set of parameters and conclusions are
sharp, even though there may be a great degree of uncertainty in those parameters.
At other times, however, this is not the case, and without better measurement of
GENERAL EQUILIBRIUM APPROACH TO BUSINESS CYCLES 191
the parameters involved, theory can only restrict the quantitative answer to a large
interval.
Findings
The final step is to report the findings. This report should include a quantitative
assessment of the precision with which the question has been answered. For the
question mentioned above, the answer is a numerical estimate of the fraction of
output variability that would have remained if variations in the growth of the
Solow residual were the only source of aggregate fluctuation. The numerical
answer to the research question, of course, is model dependent. The issue of how
confident we are in the econometric answer is a subtle one which cannot be
resolved by computing some measure of how well the model economy mimics
historical data. The degree of confidence in the answer depends on the confidence
that is placed in the economic theory being used.
The economy Lucas uses for his quantitative evaluation is very simple. There is
a representative or stand-in household and a random endowment process of the
single consumption good. The utility function is standard, namely, the expected
discounted value of a constant relative risk aversion utility function. Equilibrium
behavior is simply to consume the endowment. Lucas determines how much
192 MACROECONOMETIUCS
consumption the agent is willing to forgo each period in return for the elimination
of all fluctuations in consumption. Even with an extreme curvature parameter of
10, he finds that when the endowment process is calibrated to the U.S. consump-
tion behavior, the cost per person of business cycle fluctuations is less than one-
tenth of a per cent of per capita consumption.
This model abstracts from important features of reality. There is no investment
good, and consequently no technology to transform the date t consumption good
into the date t + 1 consumption good. As the costs of fluctuation are a function of
the variability in consumption and not in investment, abstracting from capital
accumulation is appropriate relative to the research question asked. What matters
for this evaluation is the nature of the equilibrium consumption process. Any
representative-agent economy calibrated to this process will give the same answer
to the question, so it makes sense to deal with the simplest economy whose
equilibrium consumption process is the desired one. This is what Lucas does.
Introducing the time-allocation decision between market and nonmarket activities
would change the estimate, since the agent would have the opportunity to substi-
tute between consumption and leisure. The introduction of these substitution
opportunities would result in a reduction in the estimated cost of business cycle
fluctuations as leisure moves countercyclically. But, given the small magnitude of
the cost of business cycle fluctuations, even in a world without this substitution
opportunity, and given that the introduction of this feature reduces the estimate of
this cost, there is no need for its inclusion.
In representative-agent economies, all agents are subject to the same fluctuations
in consumption. If there is heterogeneity and all idiosyncratic risk is allocated
efficiently, the results for the representative and heterogeneous agent economies
coincide. This would not be the case if agents were to smooth consumption through
the holding of liquid assets, as is the case in the permanent income theory.
Imrohoroglu (1989) examines whether the estimated costs of business cycle
fluctuations are significantly increased if, as is in fact the case, people vary their
holdings of liquid assets in order to smooth their stream of consumption. She
modifies the Lucas economy by introducing heterogeneity and by giving each
agent access to a technology that allows that agent to transform date t consumption
into date t + 1 consumption. Given that real interest rates were near zero in the
fifty-odd years from 1933 to 1988, the nature of the storage technology chosen
is that one unit of the good today can be transferred into one unit of the good
tomorrow. She calibrates the processes on individual endowments to the per-
capita consumption process, to the variability of annual income across individuals,
and to the average holdings of the liquid asset-also across individuals. For her
calibrated model economy, she finds the cost of business cycles is approximately
three times as large as that obtained in worlds with perfect insurance of idiosyn-
cratic risk. But three times a small number is still a small number.
GENERAL EQUILIBRIUM APPROACH TO BUSINESS CYCLES 193
function as a tractable way of introducing this feature. When lags on leisure are
considered, the estimate of how volatile the economy would have been if tech-
nology shocks were the only disturbance increases from 55 to near 70 per cent.
But, until there is more empirical support for this alternative preference structure,
we think estimates obtained using the economy with a time-separable utility func-
tion are better. Unlike the system-of-equations approach, the model economy that
better fits the data is not the one used. Rather, currently established theory dictates
which one is used.
Probably the most questionable assumption of this theory, given the question
addressed, is that of homogeneous workers, with the additional implication that all
variation in hours occurs in the form of changes in hours per worker. According
to aggregate data for the U.S. economy, only about one-third of the quarterly
fluctuations in hours are of this form, while the remaining two-thirds arise from
changes in the number of workers; see Kydland and Prescott (1991, Table 1).
This observation led Hansen (1985) to introduce the Rogerson (1988) labor
indivisibility construct into a business cycle model. In the Hansen world all
fluctuations in hours are in the form of employment variation. To deal with the
apparent nonconvexity arising from the assumption of indivisible labor, the prob-
lem is made convex by assuming that the commodity points are contracts in which
every agent is paid the same amount whether that agent works or not, and a lottery
randomly chooses who in fact works in every period. Hansen finds that with this
labor indivisibility his model economy fluctuates as much as did the U.S. economy.
Our view is that, with the extreme assumption of only fluctuations in employment,
Hansen overestimates the amount of aggregate fluctuations accounted for by Solow
residuals in the same way as our equally extreme assumption of only fluctuations
in hours per worker leads us to an underestimation.
In Kydland and Prescott (1991), the major improvement on the 1982 version
of the model economy is to permit variation both in the number of workers and
in the number of hours per worker. The number of hours a plant is operated in any
given period is endogenous. The model also treats labor as a quasi-fixed input
factor by assuming costs of moving people into and out of the business sector.
Thus, in this model there is what we interpret to be labor hoarding.
Without the cost of moving workers in and out of the labor force, a property
of the equilibrium turns out to be that all the hours variation is in the form of
employment change and none in hours per worker. In that respect, it is similar to
Hansen's (1985) model. For this economy with no moving costs, the estimate is
that Solow residuals account for about 90 per cent of the aggregate output vari-
ance. For this economy with moving costs, we calibrated so that the relative
variations in hours per worker and number of workers matched U.S. data. With
this degree of labor hoarding, the estimate of the fraction of the cycle accounted
for by Solow residuals is reduced to 70 per cent.
GENERAL EQUlLmRIUM APPROACH TO BUSINESS CYCLES 195
The dramatic advances in econometric methodology over the last 25 years have
made it possible to apply fully neoclassical econometrics to the study of business
cycles. Already there have been several surprising findings. Contrary to what
virtually everyone thought, including the authors of this review, technology shocks
were found to be an important contributor to business cycle fluctuations in the U.S.
postwar period.
Not all fluctuations are accounted for by technology shocks, and monetary
shocks are a leading candidate to account for a significant fraction of the
unaccounted-for aggregate fluctuations. The issue of how to incorporate monetary
and credit factors into the structure is still open, with different avenues under
exploration. When there is an established monetary theory, we are sure that general
equilibrium methods will be used econometrically to evaluate alternative monetary
and credit arrangements.
Acknowledgments
Note
* Editor's note: This chapter reprints an article originally published in the Scandinavian Journal
of Economics, vol. 93, no. 2. pp. 161-178. The authors have updated the references.
1. Koopmans subsequently became disillusioned with the system-of-equations approach. When
asked in the late 1970s by graduate students at the University of Minnesota in what direction
macroeconomics should go, Koopmans is reported by Zvi Eckstein to have said they should use the
growth model.
References
Arrow, Kenneth J., and Gerard Debreu. (1954). "Existence of an Equilibrium for a
Competitive Economy." Econometrica 22(3), 265-290.
Benhabib, Jess, Richard Rogerson, and Randall Wright. (1991). "Homework in Macro-
economics: Household Production and Aggregate Fluctuations." Journal of Political
Economy 99(6), 1166-1187.
Brunner, Karl. (1989). "The Disarray in Macroeconomics." In Forrest Capie and Geoffrey
E. Wood (eds.), Monetary Economics in the 1980s. New York: Macmillan.
GENERAL EQUll..IBRIUM APPROACH TO BUSINESS CYCLES 197
Christiano, Lawrence J., and Martin Eichenbaum. (1990). "Current Real Business Cycle
Theories and Aggregate Labor Market Fluctuations." DP 24, Institute for Empirical
Macroeconomics, Federal Reserve Bank of Minneapolis and University of Minnesota.
Cooley, Thomas F., and Gary D. Hansen. (1989). "The Inflation Tax in a Real Business
Cycle Model." American Economic Review 79(4), 733-748.
Danthine, Jean-Pierre, and John B. Donaldson. (1990). "Efficiency Wages and the Business
Cycle Puzzle." European Economic Review 34, 1275-1301.
Frisch, Ragnar. (1933). "Propagation Problems and Impulse Problems in Dynamic Eco-
nomics." In Economic Essays in Honour of Gustav Cassel. London.
Frisch, Ragnar. (1970). "Econometrics in the World of Today." In W.A. Eltis, M.F.G.
Scott, and J.N. Wolfe (eds.), Induction, Growth and Trade: Essays in Honour ofSir Roy
Harrod (pp. 152-166). Oxford: Clarendon Press.
Gornrne, Paul, and Jeremy Greenwood. (1993). "On the Cyclical Allocation of Risk."
Journal of Economic Dynamics and Control, forthcoming.
Greenwood, Jeremy, Zvi Hercowitz, and Gregory W. Huffman. (1988). "Investment, Capa-
city Utilization, and the Real Business Cycle." American Economic Review 78, 402-
417.
Harberger, Arnold C. (1962). "The Incidence of the Corporation Income Tax." Journal of
Political Economy 70(3), 215-240.
Haberler, Gottfried. (1949). "Discussion of the Econometric Approach to Business Fluctua-
tions by Tjalling C. Koopmans." American Economic Review 39, 84-88.
Hansen, Gary D. (1985). "Indivisible Labor and the Business Cycle." Journal ofMonetary
Economics 16(3), 309-327.
Hicks, John R. (1939). Value and Capital: An Inquiry into Some Furulilmental Principles
of Economic Theory. Oxford: Clarendon Press.
Hornstein, Andreas. (1993). "Monopolistic Competition, Increasing Returns to Scale, and
the Importance of Productivity Stocks." Journal of Monetary Economics 31(3), 299-
316.
Irnrohoroglu, Ayse. (1989). "Costs of Business Cycles with [ndivisibilities and Liquidity
Constraints." Journal of Political Economy 97, 1364- 1383.
Johansen, Leif. (1960). A Multi-sectoral Study of Economic Growth. Amsterdam: North-
Holland.
King, Robert G., Charles I. Plosser, and Sergio T. Rebelo. (1988). "Production, Growth and
Business Cycles: II. New Directions." Journal of Monetary Economics 21,309-341.
Koopmans, Tjalling C. (1949). "The Econometric Approach to Business Fluctuations."
American Economic Review 39,64-72.
Kydland, Finn E., and Edward C. Prescott. (1982). "Time to Build and Aggregate Fluctua-
tions." Econometrica 50, 1345-1370.
Kydland, Finn E., and Edward C. Prescott. (1990). "Business Cycles: Real Facts and a
Monetary Myth." Federal Reserve Bank of Minneapolis Quarterly Review (Spring)
3-18.
Kydland, Finn E., and Edward C. Prescott. (1991). "Hours and Employment Variation in
Business Cycle Theory." Economic Theory 1(1),63-81.
Lindahl, Erik. (1929). "Prisbildningsproblemets uppliiggning frlin kapitalteoretisk synpunkt."
198 MACROECONOMETRICS
Finn Kydland and Edward Prescott argue persuasively that the development of
dynamic, stochastic, general equilibrium theory has led to rich research possibili-
ties in macroeconomics. Their case is particularly convincing because they are
among the foremost scholars in the development and application of this research
program. The methods outlined in their review have been applied to assessing the
role of monetary and fiscal policy in business cycles, studying international cycles,
evaluating macroeconomic policies, and other questions in addition to the exam-
ples they give.
The main theme in this discussion is a simple one: conventional statistics can
be helpful in reporting the results of these modem macroeconomic investigations.
There are now a number of studies making this general point and giving details
in specific applications. Kim and Pagan (1995) provide a complete survey. I shall
illustrate some uses of statistics with an artificial example. The notation will be
simple and abstract-in the sense that I shall not refer to a specific model-but
general enough to describe some of the ideas in the research cited below.
Imagine that a model consists of specifications of preferences, budget con-
straints, technologies, and market-clearing conditions. Suppose that the necessary
(first-order) conditions can be written as
f(x" £" 0) = 0, (I)
where X, are endogenous variables, E, are shocks, and 0 are parameters. In the case
of the business-cycle model studied by Kydland and Prescott, X includes con-
sumption, output, the capital stock, and so on, £ includes a technology shock and
perhaps other shocks, and 0 includes things such as the representative agent's
intertemporal elasticity of substitution and the parameter of a Cobb-Douglas pro-
duction function.
The researcher must deduce some of the model's implications in order to an-
swer an economic question. Examples include gauging the welfare costs of busi-
ness cycles or calculating the variance of output induced by a given sequence of
shocks. If the question requires moments or sample paths of x" then the researcher
must take two steps: describe the properties of the shocks £, and solve the equa-
tions for x /.
A well-known example of describing the shocks is the stationary, first-order
autoregression for technology shocks used in many business-cycle models. Its
199
200 MACROECONOMETRICS
moments are based on those of actual Solow residuals. For future reference, sup-
pose that the shocks are modeled as
m(E" co) = 0, (2)
which is a probability law of motion describing the shocks, with parameters co.
Next, solving a model usually requires approximation and numerical methods.
Developing solution procedures has become an active area of research in macro-
economics. Approximation can begin with the first-order conditions in equation
(1) or with the optimization problem itself. Taylor and Uhlig (1990) and Ingram
(1995) review some solution methods and their properties.
Before outlining how statistics might be used to summarize the results of
experiments with the solved model, I should note that preliminary tests of the
theory usually can be undertaken, by the generalized method of moments (GMM),
without either solving the model or describing the shocks. Typically the condi-
tions in equation (1) can be summed over observations to yield moments. Those
moment conditions may be used to estimate the parameters 6. If there are more
moments than parameters (overidentification), then the ability of the same para-
meters to satisfy all the moment conditions may be used as a test of the theory.
This method was pioneered by Hansen (1982) and Hansen and Singleton (1982).
Davidson and MacKinnon (1993, ch. 17) and Ogaki (1993) outline the statistical
theory and numerous applications.
Estimation and testing by GMM has weaker requirements than does solution
of the model. The findings are consistent with a wide variety of shock processes,
which need not be specified. These weak assumptions often are enough to make
progress, for evidence against the necessary conditions in equation (1) may sug-
gest some reformulation of the theory. Nevertheless, the results also are weaker
in that this method cannot deliver predictions of future values of x or lead to
calculations of unconditional moments of x. Still GMM is an important part of the
econometrics of dynamic, general equilibrium models.
Returning to questions that do require a solution of the model, let us use the
subscript t to index observations in the model and n to index observations in
historical data. One can use actual, observed variables as shocks. For example, in
the business-cycle model described by Kydland and Prescott, one could calculate
XI from the model for an historical sequence {En} of Solow residuals. This exercise
would allow one to study the role of technology shocks (which the Solow residual
is taken to measure) by seeing what time path for output the model would
have generated under the actual shock history. This model-generated data then
could be compared with the actual history X n : n = 1, ... , N to see whether other
factors might have been important for certain elements of x. Hansen and Prescott
(1993) use this method to assess the role of technology shocks in the 1990-1991
recession.
COMMENTARY ON CHAPTER 5 201
Most studies use a different method. They begin with a statistical description
of the shocks E, parametrized so that some of its moments match those of the
historical sequence {En}. Then the research focuses on the moments of x" and a
standard example is the variance of output in the model. The moments can be
found by simulating a sequence {E,}, solving the model for a vector sequence
{x,}, and then averaging. Denote a set of such moments
r
Gr = L,g(x,; E" 8, co). (3)
,=1
Population moments, denoted G, can be approximated by making T large.
Because these studies start with this weaker information on the shocks (as
Andersen, 1991, notes), they focus on a similarly weaker property of the endog-
enous variables. The use of moments is weaker in the sense that one now cannot
ask questions such as "Does the model predict cycle turning points close to those
observed historically?" But it may also be more general in that the type of prob-
ability law used to simulate the shocks may apply over other time periods or for
other countries, where the actual sequence naturally would not.
Two types of exercises are widespread once these moments are calculated. In
type E (Experiments), described in Kydland and Prescott's review, the goal is to
calculate something in the model economy. In their examples, one calculates wel-
fare under various paths for consumption, or the variance of output under various
paths for technology shocks. Other studies are of type M (Moment-Matching), in
that they are concerned with matching moments implied by the theory with those
of actual data (GN) and perhaps with amending the basic model so as to improve
the match. At some level a type M exercise seems necessary, for confidence in
answers to questions of type E. Many macroeconomists proceed as if a model's
fit in one direction is one criterion for confidence in its answers in another. Of
course, other criteria, such as theoretical consistency, matter too, but they are not
always very restrictive empirically.
In either type of exercise, it may be important to note that the moments Gr of
x, are random variables. The randomness arises for two reasons. First, there is
sampling variability because the moments are calculated from short simulations.
Second, there may be randomness from uncertainty about the parameter values 8
and co.
Sampling variability arises because the values of the simulated, sample mo-
ments may be sensitive to a specific draw of the E, sequence. Usually, researchers
generate a small number of draws with T =N and then average the moments across
those draws, but standard statistical reasoning could improve the reporting of
results. Sampling variability in the theoretical moment,> can be removed by mak-
ing T large so that one calculates population moments.
202 MACROECONOMETRICS
many cases, the findings may not be sensitive to particular parameter settings such
as the value of the labor share. But it still may be worthwhile to show that insen-
sitivity concisely so that one has more confidence in the result.
In other cases, the properties of the theoretical model will be sensitive to the
parameter settings. A good example arises when co includes the first-order auto-
correlation ofthe technology shocks, which may not be precisely estimated. Gregory
and Smith (1990, 1993) describe some pitfalls in estimating parameters like this
by informal moment matching and by using steady-state properties. Small varia-
tions in this parameter may have significant effects on the moments G. Kim and
Pagan (1995) outline methods for sensitivity analysis. Canova (1994) has sug-
gested that parameter uncertainty can be taken into account by drawing parameters
from some distribution when calculating moments. For example, uncertainty about
a parameter value might be summarized in a uniform density across a specific
range. Moment calculations then would involve repeated draws from this density,
in addition to draws of sequences of shocks. Dejong, Ingram, and Whiteman
(1993) show how a prior distribution over parameter values 0 and co induces a
distribution over the moments G. They also show how to compare that distribution
with the estimated distribution of GN and demonstrate the comparison for a business-
cycle model. This use of priors may simplify reporting, for most researchers
currently report results for several different, fixed parameter settings.
The variability in historical, sample moments may be estimated by repeated
simulations of the theory, or it may be estimated directly from the data. Eichenbaum
(1991) gives detailed examples of this second method, in which parameters Oand
co also are estimated by GMM. He finds that the estimated proportion of historical
output volatility accounted for by technology shocks is particularly sensitive to the
uncertainty in the first-order autocorrelation and innovation variance of detrended
Solow residuals. The ratio of model output variance to historical output variance
is estimated at 0.78, in a model with indivisible labor in which these two shock
parameters are estimated. More strikingly, the standard error attached to this es-
timate is 0.64. Thus the question of what the variance of u.S. output would have
been if technology shocks were the only source of aggregate fluctuations is an-
swered imprecisely. This finding suggests that statistical evidence on experiments
is not merely quibbling, as it may show that small changes in the theory may have
large effects on the implications.
Several qualifications could be made to Eichenbaum's finding. First, Aiyagari
(1992) and Englund (1991) argue that one cannot measure the contribution of
technology shocks in a model that does not match the historical output variance.
In other words, an experiment of type E also must be of type M. If the output
variance in the model economy is less than that in the data, then a second shock
may be missing. But adding that shock will affect the model's moments and hence
conclusions about the importance of technology shocks. Perhaps because of this
204 MACROECONOMETRICS
by varying A.. The value of T could be set very large so that one uses the population
moments from the theory, removing sampling variability. And the free parameters
of the shock process «(0) and of preferences and technology (0) perhaps could be
estimated at the same time, if they are identified.
This suggestion can be thought of as a formalization of the oft-used Hodrick-
Prescott filter. In that filter the single parameter is set so that the resulting cycles
appear to resemble the predictions of the theory. For example, they retain some
autocorrelation, just as we think business cycles do. Gregory and Smith (1994)
give a several examples of detrending in this way. Because this proposal uses
some of the properties of a model to define cycles a close match between the pre-
dicted and measured properties obviously cannot be used as evidence in favor of
the theory. However, general equilibrium models, as Kydland and Prescott note,
often restrict a wide range of properties and have few free parameters so that some
moments can be used to measure cycles while others remain to be used in testing
the match between theory and evidence.
The discussion so far has outlined how statistics can be used to allow for samp-
ling variability in moments, to allow for parameter uncertainty, and to measure
cycles in a way consistent with business-cycle models. These uses of statistics
in reporting results apply even in exercises of type E in which the aim is not to
empirically match many features of the data.
In exercises of type M the aim, of course, is to reformulate the model in re-
sponse to disparities between its predictions (in directions in which it is aimed to
be informative) and some facts. It seems likely that the most interesting variations
or perturbations are those to the economics of the theory (such as those collected
by Cooley, 1994) rather than to its parameters or shocks. But research in equilib-
rium macroeconometrics typically proceeds by reformulating models in response
to some mismatch between theory and evidence.
In certain cases the use of statistics in reporting results may show that mis-
matches are not significant. For example, certain consumption-based models of
206 MACROECONOMETIUCS
asset pricing may deliver mean equity premiums that are smaller than historical
average premiums but not significantly so. Likewise, this formal reporting could
supplement subjective statements such as "The model fits surprisingly well." In
other cases this reporting style may simply strengthen results by convincing read-
ers concisely that mismatches are significant so that a reformulation is necessary.
The statistical issues may not be the most interesting or important, but that seems
all the more reason to report them so as to clear the way forward.
It is sometimes argued that formal statistical tests will reject equilibrium mod-
els because the models are simple abstractions. Practitioners have tended to es-
chew formal hypothesis tests, as Hoover (1995) has noted. However, econometric
testing can be used constructively to show in what directions a model may need
to be reformulated in order for one to have more confidence in its implications.
The test may be in a dimension with which the model is centrally concerned, so
as not to discriminate against simple, illuminating models. Moreover, there are
simple, optimizing models-such as the resilient, random-walk model of aggre-
gate consumption-that are difficult to reject with statistical tests, even though
(unlike the systems of equations approach) they are not fitted to the data as a
design criterion.
Acknowledgments
I thank the Social Sciences and Humanities Research Council of Canada and the
Foundation for Educational Exchange between Canada and the United States for
research support.
References
Aiyagari, R. (1992). "A New Method for Determining the Contribution of Technology
Shocks to Business Cycles." Working Paper 493, Research Department, Federal Re-
serve Bank of Minneapolis.
Andersen, T.M. (1991). "Comment on F.E. Kydland and E.C. Prescott, 'The Econometrics
of the General Equilibrium Approach to Business Cycles.''' Scandinavian Journal of
Economics 93, 179-184.
Canova, F. (1994). "Statistical Inference in Calibrated Models." Journal of Applied
Econometrics 9, 123-144.
Cecchetti, S.G., P.-S. Lam, and N.C. Mark. (1990). "Mean Reversion in Equilibrium Asset
Prices." American Economic Review 80, 398-418.
Cooley, T. (ed.). (1994). Frontiers of Business Cycle Research. Princeton, NJ: Princeton
University Press.
Davidson, R., and lG. MacKinnon. (1993). Estimation and Inference in Econometrics.
Oxford: Oxford University Press.
COMMENTARY ON CHAPTER 5 207
DeJong, D.N., B.F. Ingram, and C.H. Whiteman. (1993). "Beyond Calibration." Mimeo,
Department of Economics, University of Pittsburgh.
Eichenbaum, M. (1991). "Real Business-Cycle Theory: Wisdom or Whimsy?" Journal of
Economic Dynamics and Control 15,607-626.
Englund, P. (1991). "Comment on F.E. Kydland and E.C. Prescott, 'The Econometrics of
the General Equilibrium Approach to Business Cycles.''' Scandinavian Journal of
Economics 93, 175-188.
Gregory, A.W. and G.W. Smith. (1990). "Calibration as Estimation." Econometric Reviews
9(1),57-89.
Gregory, A.W., and G.W. Smith. (1991). "Calibration as Testing: Inference in Simulated
Macroeconomic Models." Journal of Business and Economic Statistics 9, 297-303.
Gregory, A.W., and G.W. Smith. (1993). "Statistical Aspects of Calibration in
Macroeconomics." In G.S. Maddala, C.R. Rao, and H.D. Vinod (eds.), Handbook of
Statistics (Vol. 11) (ch. 25. pp. 703-719). Amsterdam: North-HolIand.
Gregory, A.W., and G.W. Smith. (1994). "Measuring Business Cycles with Business Cycle
Models." Mimeo, Department of Economics, Queen's University.
Hansen, G.D., and E.C. Prescott. (1993). "Did Technology Shocks Cause the 1990-1991
Recession?" American Economic Review (P) 83, 280-286.
Hansen, L.P. (1982). "Large Sample Properties of Generalized Method of Moments Esti-
mators." Econometrica 50, 1029-1054.
Hansen, L.P., and K.J. Singleton. (1982). "Generalized Instrumental Variables Estimation
of Nonlinear Rational Expectations Models." Econometrica 50, 1269-1286.
Hoover, K.D. (1995). "Facts and Artifacts: Calibration and the Empirical Assessment of
Real-Business-Cycle Models," Oxford Economic Papers 47,24-44.
Ingram, B.F. (1995). "Recent Advances in Solving and Estimating Dynamic, Stochastic
Macroeconomic Models." Chapter 2 in this volume.
Kim, K., and A.R. Pagan. (1995). "The Econometric Analysis of Calibrated Macroeconomic
Models." In M. Wickens and M.H. Pesaran (eds.), Handbook ofApplied Econometrics.
Amsterdam: North-HolIand.
King, R.G., C.I. Plosser, and S.T. Rebelo. (l988a). "Production, Growth, and Business
Cycles I. The Basic Neoclassical ModeL" Journal ofMonetary Economics 21, 195-232.
King, R.G., C.I. Plosser, and S.T. Rebelo. (l988b). "Production, Growth, and Business
Cycles II. New Directions." Journal of Monetary Economics 21, 309-342.
Ogaki, M. (1993). "Generalized Method of Moments: Econometric Applications." In G.S.
Maddala. C.R. Rao, and H.D. Vinod (eds.), Handbook of Statistics (Vol. 11) (ch. 17).
Amsterdam: North-HolIand.
Simon, J.L. (1990). "Great and Almost-Great Magnitudes in Economics." Journal of Eco-
nomic Perspectives 4(1), 149-156.
Taylor, J.B., and H. Uhlig. (1990). "Solving Nonlinear Stochastic Growth Models: A
Comparison of Alternative Solution Methods." Journal of Business and Economic
Statistics 8, 1-17.
Watson, M.W. (1993). "Measures of Fit for Calibrated Models." Journal of Political
Economy WI, 1011-I04!.
II THE LUCAS CRITIQUE
RECONSIDERED
6 CAUSAL ORDERINGS
Stephen F. LeRoy
Introduction
211
212 MACROECONOMETRlCS
Causal Orderings
Structural Models
I distinguish variables and parameters that are structural from those that are
nonstructural. Structural variables and parameters are those that the model builder
is willing to specify to be either exogenous (determined outside the model and
subject to direct and independent intervention) or endogenous (determined by the
model and therefore not subject to direct intervention). Nonstructural variables
and parameters are those for which the model builder does not make this classi-
fication. An example of nonstructural variables would be 21 and 22 in a model
consisting of the specification that 21 and 22 are distributed as bivariate normal, and
not including any specification of which variable causes the other, or whether
instead their correlation can be ascribed to the common effect of a third variable.
A structural model-that is, a model in which at least some of the variables are
structural-is obviously the appropriate setting to analyze causal orderings.
This use of the terms exogenous and endogenous directly corresponds to their
etymology. It differs from that of econometricians, who characterize observed
variables as endogenous or exogenous according to whether or not they are cor-
related with certain unobserved variables. In general there is no connection be-
tween the two definitions.! The use here of the term structural differs from that
in some analyses, particularly when applied to parameters. In the Cowles usage
the term structural parameter means the same thing as exogenous parameter or
deep parameter here. In the present usage, by contrast, a structural parameter may
be either exogenous or endogenous, but by assumption the analyst is willing to
specify which one.
In our usage it is not necessary that all the variables and parameters in a struc-
tural model be structural. For example, in vector autoregressions the coefficient
matrices are treated as nonstructural: the analyst recognizes that the elements of
these matrices of shallow parameters are in principle functions of deep parameters
but leaves these functions unspecified. The coefficient matrices are, in effect,
treated as constants. Taking the coefficient matrices as nonstructural does not
prevent the analyst from treating the innovations as exogenous variables and the
variables determined by the VAR as endogenous variables, implying the possible
existence of causal orderings among the latter. An adjective like semistructural
might be used to describe such model, but this seems like terminological overkill.
CAUSAL ORDERINGS 213
Causal Orderings
Assume for convenience that (1) the endogenous variables form a countable set
and (2) equilibrium exists and is unique. For each endogenous variable Yn there
exists a vector of exogenous variables e(Yn) consisting of the smallest set needed
to determine Yn. The vector e(Yn) will be termed the exogenous set for Yn. 2 These
exogenous sets will playa major role in the analysis of causal orderings.
Then the solution to any model consists of a sequence of functions In:
n = 1,2, ....
The notation YI ~ Y2 is defined to mean that YI causes Y2- 3 We have Yt ~ Y2 if three
conditions-the strict inclusion condition, the composition condition, and the
nonconstancy condition-are satisfied. 4 The strict inclusion condition is satisfied
if e(YI) CC e(Y2) (e(YI) is a proper subset of e(Y2». If the strict inclusion con-
dition is satisfied, there exists at least one exogenous variable that affects Y2 but
does not feed back onto Yl' guaranteeing the asymmetry of causations.
Define Z2-1 as the vector of exogenous variables that appear in e(Y2) but not in
e(Yt); satisfaction of the strict inclusion condition guarantees that Z2-1 is nonempty.
The composition condition consists of the requirement that the equation express-
ing the solution for Y2'
Y2 = h(e(Y2»,
(6.1)
(6.2)
(6.3)
Equation (6.2) says that the number of nematodes on a farmer's land is detennined
by the amount of pesticides he applies and rainfall, while equation (6.3) says that
the amount of pesticides the farmer applies is a function of nematodes and the cost
of pesticides.sHere n and p are endogenous, while r and c are exogenous. For the
present the parameters a; and Pi are treated as nonstructural-that is, as constants.
°
In the special case PI = the model has the causation p ~ n. Failing the restriction
PI = 0, the question "What is the effect of pesticides on nematodes?" is seen to
be ambiguous: pesticide use is affected by both rainfall and pesticide cost, and
the effect on nematodes depends on which caused the change in pesticide use
214 MACROECONOMETIUCS
(6.6)
where the x's are exogenous variables and the y's are endogenous variables. This
model has Yl ~ Y2 and Yl ~ Y3: an intervention on the exogenous variables in
e(YI)-xlOand XII-affects Y2 and Y3 only through its effect on YI' Therefore the
questions "What is the effect of YI on Y2?" and "What is the effect of YI on Y3?"
are well posed: the questions presuppose an intervention on XIO and XII with all
other exogenous variables held constant, and all interventions on XII and XIO result-
ing in a particular change in YI have the same effect on Y2 and YJ.
The model has no other endogenous variables that are causally ordered. In
particular, it does not have Y2 ~ Y3 because of failure of the composition condition.
Therefore the question "What is the effect of Y2 on Y3?" is not well posed: char-
acterizing an intervention only by its effect on Y2 does not give enough information
about the intervention to determine its effect on YJ. On the contrary, of the vari-
ables in e(y2)-xlO' XII' X20' X21' and xn-the first two also affect Y3 through their
effect on YI-
Informally, it is obvious that exogenous variables and the endogenous variables
CAUSAL ORDERINGS 215
that depend on them are also causally ordered: there is no ambiguity about the
intervention if the variable being intervened on is itself exogenous. To take formal
account of this fact we will extend the definition of causal orderings so as to have
Xi ~ Yj under exactly the same conditions as are required for Yi ~ Yj' with the
specification that the exogenous set of an exogenous variable consists of that
variable alone. Thus extended, for any model causality is a relation on the union
of the set of the names of the exogenous and that of the names of the endogenous
variables. It is easy to verify that the relation ~ is irreftexive, asymmetric, and
transitive, thus constituting a strict partial order, with the exogenous variables as
minimal elements.
Extensions
CondftionalCausalOrderings
Suppose that the endogenous variables in the nonsingleton set b jointly cause Y2
and that Yl E b.Is there any sense in which YI by itself can be said to cause Y2?
With b ~ Y2 there potentially is associated the conditional causation Yl ~ Y2 I
(b - Yl). This definition and notation indicate that the intervention on e(YI) is
understood to be restricted so that the variables in b - YI (the elements of bother
216 MACROECONOMETRICS
than Y.) are held constant. Under the restriction on the domain of f. the strict
inclusion and composition conditions for Y. ~ Y2 are satisfied by definition. If
under the domain restriction the nonconstancy condition is also satisfied, then the
conditional causation obtains.
For example, consider the model
(6.7)
(6.8)
(6.9)
This model has the joint causation (Y., Y2) ~ Y3 but has no endogenous variables
that are simply causally ordered. It also has the conditional causations Y. ~ Y31 Y2
and Y2 ~ Y31 YI' The first of these has the interpretation that all interventions on
e(YI) that are associated with a particular change in Yl and with no change in Y2
map onto a particular value of Y3-
The idea of conditional causal orderings has a serious drawback that impairs
its usefulness in applications (in fact, I engage in the present discussion of con-
ditional causal orderings only because having access to the concept of conditional
causal orderings allows a compact statement below of the Lucas critique): postu-
lating across-variable restrictions on the hypothesized interventions on exogenous
variables implicitly presupposes the existence of functional relations among the
exogenous variables. This is inconsistent with their exogeneity. Thus the idea of
conditional causations conflicts with the attribute of exogenous variables that they
vary independently.
There is also the problem that the existence of joint causations does not nec-
essarily imply the existence of conditional causations. The restriction on the do-
main off. may limit Y. to a single value. In that case g can be chosen to be a trivial
constant function of YI' so there is no conditional causation. For example, consider
the model
where the a's are constants. This model differs from the model (6.7) through (6.9)
in that the coefficients of the y's and one of the intercepts are constants rather than
exogenous variables. Here the exogenous sets for Y. and Y2 each consist of the
single exogenous variable x 10' In this model there exists the joint causation (YI' Y2)
~ Y3 but not the conditional causations YI ~ Y3 I Y2 or Y2 ~ Y3 I YI' The reason is
that, in the first case, if Y2 is held constant a unique value for YI is implied, so g
can be written as a constant function of YI' If there exist conditional causations
CAUSAL ORDERINGS 217
associated with a given joint causation, the joint causation is separable, following
the terminology of Hicks's valuable (1979) discussion.
In models that are linear in variables, with every causation YI ~ Y2 there is asso-
ciated a coefficient giving the effect of a unit change of YJ on Y2' This coefficient
is a parameter if it is structural or a constant if it is nonstructural. Having specified
nothing about which variables are observable or about the probability distribution
of unobserved exogenous variables, there is no presumption that the coefficients
associated with causations are identified. However, there is an important special
case in which the coefficients associated with causations are identified and easily
estimated. Suppose that all exogenous variables are independently distributed
random variables. Such a model will be called complete because it gives a com-
plete account of the correlations among endogenous variables, assigning no role
to uninterpreted correlations among exogenous variables. If a linear model with
causation YI ~ Y2 is complete and if YI is observable, then the assumptions of
the Gauss-Markov theorem are met and unbiased and consistent estimates of the
218 MACROECONOMET~CS
coefficient associated with the causation can be obtained using ordinary least
squares.
The contrapositive version of this result provides a convenient way of proving
that two variables are not causally ordered, particularly in the hands of those
familiar with econometric estimation theory. If in a complete model one knows
that a particular regression coefficient is estimated with bias or inconsistency
under ordinary least squares, it cannot be the case that the independent and
dependent variables of the regression are causally ordered, with the parameter in
question being that associated with the causation. Knowledge that estimation
problems occur therefore indicates absence of causation.
For example, in the nematodes-pesticide example, if /31 :¢ 0, an OLS regression
of nematodes on pesticides will result in an inconsistent estimate of ai' even if r
and c are independently distributed so that the model is complete. Therefore we
cannot have the causation p => n, with a I as the associated parameter. However,
under the restriction /31 =0 simultaneous equations bias is eliminated and a l is
estimated consistently by OLS. This fact suggests that we may have p => n with
associated parameter a I; independent investigation is required to confirm this (and
does confirm it, as was seen above).
For another example, consider the model
x, = AX'_1 + U"
U, = PUt_I + e"
with the initial values X o and U o and the e, as exogenous variables. Suppose also
that the model is complete, so that x o' uo' and the e, are distributed independently.
Econometricians know that if A. is estimated by an ordinary least-square regression
of x, on X'_I' bias will result in small samples because of the correlation between
u, andx'_l induced by the autocorrelation of the error. Applying the above results,
an econometrician could conclude immediately that the causation xt _ 1 => x, does
not hold without bothering to verify that the formal conditions for a causal order-
ing are not satisfied.
Simon
Leamer
Leamer (1985, pp. 262 ff.) presented alternative definitions of terms correspond-
ing to causal orderings as defined here:
Structural: A parameter is structural if it is invariant under any modification of the
system selected from a specified family of modifications. to
Exogenous: If the observed conditional distribution of the variable Ygiven a set
of variables x is invariant under any modification of the system selected from a
specified family of modifications that alter the process generating x, then the
variables x are said to be exogenous to y.
What is noteworthy about Leamer's definitions is that they treat parameters and
variables separately: the term structural is used for parameters, whereas exogenous
is used for variables. Further, the definitions themselves differ: structuralness is
directly related to invariance of parameters under intervention, whereas exogeneity
is defined in terms of invariance of conditional distributions. In the present paper,
in contrast, the characterization of causal orderings does not depend on any dis-
tinction between parameters and variables-for the purpose of defining causal
orderings it is immaterial whether some or all of the variables of a system are
treated as terms of families (that is, as elements of indexed sets).
It does not appear that the distinction between variables and parameters is used
in any consistent sense in the economics literature: it is easy to find exceptions for
the possible bases for such a distinction that come first to mind. For example,
because Leamer's definition refers to probability distributions in connection with
variables but not parameters, one might define variables as random and parameters
220 MACROECONOMETRICS
as nonrandom. But that will not do: learning models have random parameters and
all dynamic deterministic models have variables that are nonrandom. Alterna-
tively. parameters might be defined as constant functions on an index set that is
interpreted as time. This characterization will not do either: the capital-output ratio
in a Solow growth model is constant in steady-state equilibrium, yet one would not
call it a parameter. Similarly. in game-theoretic models with imperfect informa-
tion it is common knowledge that players' types are unchanging over time, yet
types would normally be classified as variables. not parameters. Finally, the term
parameter might be used to connote structuralness or exogeneity. However, then
the term structural parameter would contain a redundancy, and shallow para-
meter (as opposed to deep parameter) a contradiction.
It follows from this, not that the distinction between parameters and variables
is meaningless but that it acquires different meanings in different contexts. For
example, we often describe a model as linear, meaning linear in variables, with the
coefficients being parameters, not variables. Here the distinction between para-
meters and variables derives from the assumption of linearity. For another example,
we specify a model as incorporating rational expectations, meaning that agents are
assumed to know all parameter values but not necessarily all variable values. For
the meanings of linear in variables and rational expectations to be unambiguous,
it is essential that the distinction between variables and parameters be specified
clearly, but the distinction itself does not necessarily correspond to some preex-
isting difference between parameters and variables. The notation adopted in set-
ting out a model is chosen to delineate the distinction between parameters and
variables, usually without explicit discussion. The arbitrariness of the parameter-
variable distinction means that it is inappropriate to distinguish, as Leamer did,
structuralness of parameters from exogeneity of variables, at least in the absence
of a clear characterization of the difference between parameters and variables and
an explanation for the differing definitions of structuralness and exogeneity in
terms of this characterization.
Rather than take the parameter-variable distinction as logically prior to the
definition of causal orderings, it appears to make sense instead to use some of the
conceptual apparatus of causal orderings to characterize the parameter-variable dis-
tinction. By calling something a parameter we (often) are implicitly committing to
the assumption that its exogenous set consists only of exogenous parameters, not
exogenous variables. Thus the definition of parameter would imply that an inter-
vention on variables alone leaves endogenous parameters unaffected. This charac-
terization of parameters corresponds to a parallel characterization of variables: by
labeling a variable y, we often are committing to the assumption that the exogenous
set for that variable consists only of exogenous parameters and exogenous variables
dated up to and including date t: exogenous variables dated after t cannot influence
Yr' Like the alternative characterizations of the parameter-variable distinction
CAUSAL ORDERINGS 221
Applications
Vector Autoregressions
as in Jacobs, Leamer, and Ward (1979) and Cooley and LeRoy (1985). Here the
exogenous variables mo, Yo, Ell' and Ez, are assumed to be contemporaneously and
serially uncorrelated, so that the model is complete in the sense defined above.
First, we have the obvious point (obvious because El , and Ez, are directly speci-
fied to be exogenous variables) that the causations Cl' => m" El, => y" Ez, => m"
and Ez, => y, obtain, with coefficients 1/(1 - Or), rl(l - Or), O/(l - Or) and 1/
(I - Or», respectively. These coefficients are unidentified. Restrictions like = °
o (so that realizations of Ez, do not feed back to m,) or r + f3J.1 = 0 (so that the
steady-state effect of money on income is zero), however, identify all coefficients.
VARs are usually written in the triangular form
m, = Pllm,-1 + PI2Y,-1 + 11mI' (6.12)
with equations (6.12) and (6.13), so that 8= y, Pij = f3ij (i,j = 1, 2), and in particular
11mt and 11yt equal Elt and ti" respectively. Therefore the causations 11m' ~ m,. 11m'
~ Y" 11Yt ~ Y, (but not 11Y, ~ m t) now obtain. Further, under the restriction 8 =
o the causations 11mt ~ m,+j' 11m' ~ Yt+j' 11yt ~ m'+j and 11yr ~ Yt+j (j ~ 1) also
obtain, with associated coefficients that are just identified.
The causation m, ~ Yt obtains only under extremely restrictive conditions;
surprisingly, even the strong restriction 8= f312 = 0, which assures that tit does not
feed back onto either current or future values of m" does not imply m, ~ Y, .
Therefore the question "What is the effect of money on income?" is ambiguous
even under 8= f312 = O. However, the joint causation (y" m,. m'+I' ... , m t+n ~ Yt+n)
obtains under the restriction 8 = f312 = O. This fact is useful in interpreting the
"dynamic simulations" of Keynesian policy analysis, according to which the
effect of changing the assumed future trajectory of some policy variable like
the money stock is calculated by solving a macroeconomic model forward in time
twice, once using the reference path for the policy variable and once using a
perturbed path. We see that under the parameter restriction 8 = f312 = 0 the "dy-
namic multipliers" dYt+i+/dmt+i (i, j ~ 0) calculated from the exercise just de-
scribed collectively give the effect of an alteration in the trajectory for the money
stock on the trajectory for income. However, note that because 8 = f312 =0 is not
sufficient for the simple causation m,+i ~ Yt+i+j' the dynamic multipliers individu-
ally cannot be interpreted as giving the effect of mt+i on Y,+i+j'
It remains to ascertain the relation between Granger causation and causal
orderings as defined in this paper. In the above model money is exogenous in the
Granger sense if it is equally well predicted by its own lagged values as by these
plus lagged values of income. This property implies and is implied by the para-
meter restriction 8f322 + f312 = O. It is seen that Granger noncausation of money
is neither necessary nor sufficient for the causation 11m, ~ Yml' since 8f322 + f312 =
o neither implies nor is implied by 8 = O. However, rejection of Granger non-
causation does imply rejection of 8 = f312 = 0, implying that the dynamic multi-
pliers described in the preceding paragraph cannot be interpreted as corresponding
to a joint causation.
e
Let denote the set of names of the deep parameters and 'l'the set of the names
of the shallow parameters. Then a full structural model (expressed in tenns of its
solution, as above) consists of
If'; = /;(e(lf';»,
Yi = gj(e(y;),
for", E '1'. Here e( 'IIi) c e, while e(Yi) C (e U X), where X is the set consisting
e
of the names of the exogenous variables. The fact that e( "';> is a subset of alone
rather than a general subset of e U X reflects the convention, discussed in the
fourth section, that interventions on exogenous variables do not affect endogenous
parameters.
As noted above, the argument just made implies that full structural models may
have the joint causation A ~ Yi for each Yi' where A is some subset of the shallow
parameters '1'. Suppose now that one treats the elements of A as exogenous vari-
ables and conducts interventions on them independently. Analytically this corre-
sponds to treating a joint causation as a conditional causation. However, if there
are fewer deep parameters than shallow parameters, as is virtually always the case,
the requisite separability property does not obtain (see the third section), so the
indicated conditional causation does not exist and the exercise is inadmissible. We
have here one version of the Lucas (1976) critique: according to Lucas, Keynesian
macroeconomists conducted policy analysis by treating the coefficients of linear
equations as exogenous variables subject to independent intervention. In full struc-
tural models, however, these coefficients are shallow parameters that are linked by
across-equation restrictions to the deep parameters of preferences and technology.
Because there are few deep parameters and many shallow parameters, it is seldom
possible to interpret the exercise of altering one shallow parameter and holding all
others constant in tenns of an intervention on the deep parameters.
The Lucas critique has persuaded many economists that policy evaluation can be
conducted correctly only using full structural models: "Disentangling the para-
meters governing the stochastic processes that agents face from the parameters of
their objective functions would enable the econometrician to predict how agents'
decision rules would change across alterations in their stochastic environment.
Accomplishing this task is an absolute prerequisite of reliable econometric policy
evaluation" (Hansen and Sargent, 1981). However, contrary to the instruction of
the Lucas critique, it is not the case that hypothetical alterations in agents' stochastic
environments can be properly modeled only via parameter interventions (that is,
CAUSAL ORDERINGS 225
Conclusion
Notes
I. However, in models that are complete in the sense defined below the two definitions are closely
related.
2. It is evident that exogenous sets are unique. Suppose, on the contrary, that e(y.) and e'(y.) are
exogenous sets for y•. If these are distinct, for some exogenous variable x we have x E (Y.) but x e
e'(y.). But this cannot occur: if the variables e'(y.) are held fixed, then Y. is fixed, implying that
interventions on x do not affect Y•. However, then e(y.) contains a redundant element. contrary to the
definition of exogenous sets.
3. Here we might instead read y, ~ Y2 as "y, is a cause of y/' to emphasize that endogenous
variables other than y, may also be causes of Y2- The chosen reading has the advantage of brevity.
4. Under the characterization here, the restrictions on the solution functions f. implied by the
existence of causal orderings are identical to those studied in functional structure theory (see, for
example, Blackorby, Primont, and Russell, 1978). Therefore the formal theory of causal orderings is
closely linked to such topics as demand theory and the theory of aggregation.
5. This example is drawn from the SASIITS User's Guide (Version 5 Ed.), p. 472. I am indebted
to Mark Fisher for the reference.
6. Formally, joint causation is an ordering, not on the set of variable names but on the power set
of that set.
7. Note that the term variable is used in different senses in this sentence and in the first sentence
of this subsection. In the first sentence variables are distinguished from constants (constants are
standins for particular numbers, whereas variables are interrelated via the functions of the model),
while in this sentence they are distinguished from parameters, where the latter are not dated. Rather
than weigh down the discussion by adopting special terminology that distinguishes the two senses of
variable, I will rely on the context to make the meaning clear.
8. Here g-'(<P2, Z2-l) denotes (y, E Y, such that g(y" Z2-') E <P2}.
CAUSAL ORDERINGS 227
9. In fact, changing Z2-1 may result in g(YI' Z2-I) becoming a constant function ofYl' invalidating
the causation. This occurred in the firing squad example considered above: if all other members of the
firing squad missed, then the event that member i hit the victim caused the event that the victim was
killed. On the other hand, if any of the other members of the firing squad also hit the victim. the event
that member i's bullet hit the victim did not cause the death of the victim. other things equal.
10. Leamer used the term modification with exactly the same meaning as intervention has here.
References
Blackorby, Charles, Daniel Primont, and R. Robert Russell. (1978). Duality, Separability
and Functional Structure: Theory and Applications. New York: North-Holland.
Cooley, Thomas F., and Stephen F. LeRoy. (1985). "Atheoretical Macroeconometrics: A
Critique." Journal of Monetary Economics 16,283-308.
Hansen, Lars Peter. and Thomas J. Sargent. (1981). "Fonnulating and Estimating Dynamic
Linear Rational Expectations Models." In Robert E. Lucas, Jr. and Thomas Sargent
(eds.), Rational Expectations and Econometric Practice. Minneapolis: University of
Minnesota Press.
Hicks, John. (1979). Causality in Economics. New York: Basic Books.
Jacobs, Rodney L., Edward E. Leamer, and Michael P. Ward. (1979). "Difficulties with
Testing for Causation." Economic Inquiry 17,401-413.
Leamer, Edward E. (1985). "Vector Autoregressions for Causal Inference." Camegie-
Rochester Conference Series on Public Policy.
LeRoy, Stephen F. (1993). "On Policy Regimes." Chapter 7 in this volume.
Lucas, Robert E., Jr. "Econometric Policy Evaluation: A Critique" (1976). In K. Brunner
and A. Meltzer (eds.), The Philips Curve and Labor Markets. Carnegie-Rochester Series
on Public Policy (Vol. 1). Amsterdam: North-Holland.
Simon, Herbert A. (1953). "Causal Ordering and Identifiability."1n Wiliam C. Hood and
Tjalling C. Koopmans (eds.), Studies in Econometric Method. New Haven: Cowles
Foundation.
Commentary on Chapter 6
Clive W.J. Granger
Discussions of causality have been occurring for at least three thousand years, and
the fact that they continue suggests both that the topic is a difficult one and that
there is a continuing real demand for an acceptable definition. Anyone can propose
their own definition of causality by emphasizing some aspect of causality that is
considered to be of particular or obvious importance. For example, Hoover and
Sheffrin (1992) equated causality with controllability and base a definition on that,
whereas the definition that I have marketed and defended-for example, in Granger
(l980)-emphasizes that the cause may occur temporally before the effect and
contain special information about it, so that necessarily the cause can help forecast
the effect.
Each definition, when considered formally, will be based on a number of assump-
tions, which mayor may not be acceptable to potential users. From the definition
there will be a number of consequences, which mayor may not be thought to be
sensible. It is virtually impossible to prove that one definition is inherently superior
to another using just logical reasoning as, starting with the belief that one defini-
tion is correct, including that all its assumptions are true, will probably lead to an
alternative definition producing "contradictory" results, or even results that seem
lacking in common sense. The various definitions do not encompass each other,
and one cannot be nested in another.
For leRoy causality one has to be able to state a set of variables, denoted e(y),
which completely determine the variable y. If y is a random variable, the condi-
tional distribution F(y I e (y» is singular, so that y = f(e (y» for some function of
f, without any error or uncertainty. It is obviously an act of faith that such a set
e(y) exists, and that it is not uncountably infinite. To see the problem, suppose I
note that at the age of fifty-nine I am six feet tall. I would need to list all of the
reasons for or determinants of that fact. However, in the definition being consid-
ered a number of helpful assumptions are made, including that e (y) consists ofjust
a countable set of variables, that equilibrium exists and is unique and that some
model is available that provides the function f
Now suppose that the vector of variables under consideration can be decom-
posed into two distinct types, called endogenous and exogenous. This decom-
position is performed by the researcher, and the two types of variables are not
carefully defined, although the decomposition is very important for the paper. It
is said that an exogenous variable is "determined outside the model and subject
229
230 MACROECONOMETIUCS
quantum physics being widely accepted, and now they generally admit that cau-
sality can often be stochastic. If you smoke, you do not necessarily get cancer, but
you do increase the probability of getting the disease. For a full discussion see
Skynns and Harper (1988). This has proved to be such a useful generalization, it
would be a clear backward step to determinism, in my opinion, which follows
from the assumption that all exogenous variables can be defined and listed.
One of the most curious and certainly controversial aspects of the new causal
approach follows from the model given in the third section:
where "xo and U o and the e, are exogenous variables" and "the model is complete,
so that Xo ' UO ' and the e, are distributed independently." It is pointed out that if A.
is estimated by OLS it will usually be biased in small samples and so "XI-1 cannot
cause x,," This statement is made regardless of the true value of p. Is it true of p
= =
0, for example, or if p I? Is the bias result true if I p I is less than one but the
true A. is larger than one? Why should the existence of causation depend on the
method of estimation used for a parameter: if a method can be found that produced
an unbiased estimate of A., is there then causation? Suppose that there is an esti-
mator that produces a positive bias in the estimate of A. for sample sizes up to no,
a negative bias for sizes over no, but no bias for a sample size exactly no. Can it
really be possible that X,_I causes x, only for one particular sample size and no
other? This is surely a confused argument between causal variables and unknown
parameters. Most definitions of a deep concept such as causality would be based
on the properties of populations, not on estimated models. The tests of causality
may have to be based on data considerations, but one does not need to mix the
question of definition with that of testing and thus of the estimation of models. The
model given above can also be used to indicate a further weakness of the LeRoy
definition, in that everything is conditional on the particular infonnation set being
considered. Thus, for example, the independence of x o , uo , and the E, may if one
is just considering the x, series and how it is generated but may not be true if it
is embedded within a larger vector of variables. It is well known that a pair of
Gaussian random variables x, y can be uncorrelated, and thus independent, within
the set (x, y), but then their partial correlation corr (x, y I z) need not be zero, so
that they are not independent in the set (x, y, z). Independence is not a pervasive
property, so that once found it is always present, but rather it depends on the
context.
It seems that the new definition runs into logical problems when it tries to
encompass all aspects of model generation, specification, and estimation, particu-
larly when the models are dynamic. There is one area in which it can have a very
232 MACROECONOMETRICS
References
Engle, R.F., D. Hendry, and J-F. Richard. (1983). "Exogeneity." Econometrica 51, 277-
304.
Glymour, c., R. Scheimes, P. Spines, and K. Kelley. (1987). Discovering Causal Structure
San Diego: Academic Press.
Glymour, c., and P. Spines. (1988). "Latent Variables, Causal Models and Overidentifying
Constraints." Journal of Econometrics 39.
Granger, C.W.J. (1980). "Testing for causality: A Personal Viewpoint." Journal of Eco-
nomic Dynamics and Control 2, 329-352.
COMMENTARYONCHA~R6 233
Hoover, K.D., and S.M. Sheffrin. (1992). "Causation, Spending and Taxes: Sand in the
Sand-box or Tax Collector for the Welfare State?" American Economic Review 82,
225-248.
Pearl, J., and T.S. Venna. (1991). "A Theory of Inferred Causation." In J.A. Allen, R.
Filkes, and B. Sandewall (eds.), Principles ofKnowlege and Reasoning: Proceedings of
the Second International Conference. (San Mateo, CA: Morgan Kaufman).
Skynns B., and W.L. Harper. (1988). Causation, Chance and Credence Dorchrecht: Kluwer.
7 ON POLICY REGIMES
Stephen F. LeRoy
Introduction 1
235
236 MACROECONOMETRICS
The Theory of Economic Policy that Lucas criticized originated with Tinbergen
(1952) and was developed by Cowles Commission economists (Simon, Marschak,
and Klein) in the 1950s. The Cowles group contributed a variety of extremely im-
portant developments to economic and econometric theory: the theories of causal
orderings, of econometric identification, and of simultaneous equations bias all
were developed within the same broad project as the Keynesian mode of policy
analysis. General equilibrium theorists of the stature of Debreu, Hurwicz, and
Koopmans worked on related Cowles projects. Given the close association be-
tween macroeconomists and general equilibrium theorists, it is surprising to learn
that the two traditions are in fundamental conflict.
In fact, there is no inconsistency between the Theory of Economic Policy as
formulated by the Cowles group and general equilibrium theory. In this paper it
will be argued that the Theory of Economic Policy as Lucas characterized and
criticized it is substantially correct; the extension of the Cowles analysis to rational
expectations does not alter the original analysis at any deep level. Rather, the
conflict is between Lucas's recommended mode of policy analysis and a consist-
ent application of rational expectations. I will argue that Lucas's preferred formu-
lation of policy analysis via parameter interventions amounts to replacing the
assumption of rational expectations with that of stationary expectations.
The Lucas critique is reviewed in the second section. In the third section the
Lucas critique is contrasted with the received Cowles approach. The theme is that
the two are diametrically opposed and that new classical economists confuse
rational expectations with stationary expectations. Three examples are discussed
in the fourth section. The discussion then deals with the consequences for such
new classical doctrines as policy ineffectiveness of maintaining a consistent frame-
work of rational expectations, and the sixth section concludes.
Lucas took the model characterizing the effect of policy variables on the economy
to be
computed as the solution to the optimization problem provides the answer to the
policy analysis question.
Lucas's criticism was that if the x, are subject to change in the future but are
not characterized probabilistically, then agents cannot be represented as maximiz-
ing utilities subject to well-defined constraints. Further, there can be no presump-
tion that the parameters 9 will be invariant to the interventions in the X" as assumed
in the optimal control exercise.
Lucas's counsel to economists as model builders was that they represent policy
choice by hypothetical parameter changes. This instruction had as its counterpart
his injunction to economists as advice givers that they restrict themselves to ad-
vocating that policy be conducted according to simple rules. Lucas recognized that
the transition to a regime of fixed rules can be done in many ways and that the
argument for fixed rules gives no guidance in choosing among these. It appears to
follow that economists have nothing to contribute by way of advice to policy
makers in any setting in which the past conduct of policy is taken as given. Lucas
(1980 p. 258) explicitly accepted this implication:
To one with some responsibility for monetary policy in 1974, say, it is not very helpful
to observe that monetary growth "should have" proceeded at a constant 4 percent rate
for the 25 years preceding.... What advice, then, do advocates of rules have to offer
with respect to the policy decisions before us right now?
This question does have a practical, men-of-affairs ring to it, but to my ears, this ring
is entirely false. It is a king-for-a-day question which has no real-world counterpart in
the decision problems actually faced by economic advisors....
Economists who pose this "What is to be done, today?" question as though it were
somehow the acid test of economic competence are culture-bound (or institution-bound)
to an extent they are probably not aware of. They are accepting as given the entirely
unproved hpyothesis that the fine-tuning exercise called for by the Employment Act [of
1946] is a desirable and feasible one.
I noted in the previous section that Lucas made two points in his critique-that
policy changes are to be modeled as parameter shifts and that policy should be
governed by simple rules. The points are distinct and arguably unrelated. Even the
intended audiences are different: the first is addressed to economists and concerns
how to model policy; the second is addressed to government and concerns how to
conduct policy. Yet Lucas argued them simultaneously. For example, consider his
discussion of what happens when agents are confronted with arbitrary changes
in their environment. This question came up twice in Lucas's paper: once in the
context of his criticism of Keynesian policy evaluation and once in the context of
his discussion of regime shifts under his preferred formalization of policy analysis.
The essence of Lucas's criticism of Keynesian policy evaluation was that because
240 MACROECONOMETRICS
The two points Lucas made-that policy changes should be modeled as parameter
shifts rather than variable shifts and that policy should be governed by simple
rules-are distinct and should be appraised separately. I take them in order. It is
difficult to understand the basis for Lucas's distinction between policy evaluation
defined as selection of an arbitrary sequence for a forcing variable, as under the
theory of economic policy, versus policy evaluation as selection of a policy re-
sponse function, as in his recommended mode of analysis. I have already observed
ON POLICY REGIMES 241
Simple Rules
As has been seen, one of Lucas's major purposes, both in the "Critique" and else-
where, was to argue that policy should be guided by preannounced simple rules,
as opposed to either discretion or policy rules involving feedback. In the preferred
new classical models of macroeconomic activity-those expressing equilibrium
GNP as a function of the unexpected component of money or prices-the reduced
form for GNP does not depend on the parameters of the policy rule. This is so
because an intervention on the feedback parameters of the money supply process
merely shifts the distribution of the money stock, implying that the distribution of
the unexpected component of money is unaffected (see the discussion of Sargent-
Wallace, 1976, below). Thus simple rules do just as well as feedback control rules.
In light of the discussion of the second and third sections, one may distinguish
two versions of the foregoing argument for simple rules: the rational expectations
242 MACROECONOMETRICS
version and the stationary expectations version. The rational expectations version
argues that if agents in one country know that macroeconomic policy has been, is
now, and always will be governed by a feedback control rule, and if in another
otherwise identical country agents know that policy has been, is now and always
will be governed by a simple rule, the two countries will have similar economic
performances. This argument is correct, given the validity of the assumed model.
In most discussions, however, the context of the discussion concerns real-time
policy changes,2 not hypothetical timeless comparisons, and the argument asserts
that a fully understood change from one policy rule to another will not affect
economic performance (again, see the discussion of Sargent-Wallace, 1976, be-
low). Again the argument is correct, but it must be understood that now the setting
is one of stationary expectations, not rational expectations: under both the old and
new control rules agents are modeled as always being certain that the policy
parameter will continue at its current value even though it changes over time as
the policy regime shifts.3
Examples
We see that even though the Lucas critique is usually read as a criticism of
Keynesians for not assuming rational expectations, in fact it is Lucas's own con-
cept of policy regimes that is vulnerable to this criticism (if, indeed, it is a criti-
cism). It is interesting to trace out this interplay in Lucas's paper. To do so I
reconsider two of the examples discussed by Lucas, and also analyze the 1976
model of Sargent and Wallace.
Lucas's first example was based on the permanent income hypothesis of Friedman
(1957). As is well known, the permanent-income hypothesis explains the slope of
the observed consumption function in terms of the variances of the permanent and
transitory components of income. Accordingly, Lucas observed, the effects of a
policy consisting of a sequence of income supplements depend critically on whether
the supplements are perceived as permanent or transitory. The Keynesian analyti-
cal practice, however, was the same for either case: simply add the supplement to
the forecast of income and predict consumption by applying the estimated con-
sumption function.
An analyst wishing to determine the effect of income supplements within the
framework of the permanent income hypothesis would surely recognize that
the permanent-income hypothesis specifies that agents can distinguish between
ON POLICY REGIMES 243
In his second example Lucas criticized Hall and Jorgenson's (1967) analysis of the
effect of an investment tax credit on the level of investment. Hall and Jorgenson
had treated the tax credit as a parameter, implying that any tax rate change is re-
garded as permanent. Lucas pointed out that if the tax rate is being used as a tool
of countercyclical stabilization policy, however, then tax-rate changes are obvi-
ously not permanent. Lucas proposed modeling the tax rate not as a parameter but
as a realization from a Markov process with transition probabilities approximating
the real-world frequencies of tax rate changes.
The conclusion Lucas drew at the end of his section on Hall-Jorgenson was that
the example illustrated the importance of modeling policy as generated by a fixed
rule-in this case the stochastic process he proposed as a model for the tax rate-
rather than characterizing policy by an arbitrary parameter. Correspondingly, Lucas
argued by example, a change in tax policy should be represented as an altered
realization of the random variable the probability distribution of which this policy
rule consists. Yet in his conclusion at the end of the paper Lucas advocated model-
ing policy change as regime choice, precisely the opposite point. It was Hall and
Jorgenson who modeled policy change as regime choice and Lucas who, in the
Hall-Jorgenson section of his paper, proposed modeling a change in tax policy as
an altered realization of a random variable. Hall and Jorgenson's analysis of tax
rate changes appears to conform to the pattern advocated by Lucas in his con-
clusion; Lucas, by representing policy change as an altered realization of a vari-
able, appears in his reformulation of Hall and Jorgenson to advocate doing policy
analysis in exactly the way he criticized in his introduction and conclusion.
244 MACROECONOMETRICS
Lucas's third example dealt with the Phillips curve. It is essentially the same
as the Sargent-Wallace model, to which I now tum.
Sargent and Wallace's (1976) paper presented the argument for rules in the
context of a linear-quadratic new classical macroeconomic model. Sargent and
Wallace's purpose was to show that the general presumption in favor offeedback
control implied by optimal control theory depends on the assumption that the
structure generating GNP is constant over time as control rules change. If, however,
as rational expectations appeared to them to imply, the structure changes as con-
trol rules change, then the presumption in favor of feedback control disappears.
Sargent and Wallace assumed that GNP Y, is related to the money stock m,
according t04
Yr = AYt-I + 13m, + u, (7.3)
and that the goal of policy is to minimize the variance of Yr' They began by
reviewing the case for feedback control. The feedback control rule
m, = gY'_1 (7.4)
results in
2
V()- (1u
Y, - 1 - (A + pg)2 '
which, with A*- 0, is strictly lower under the optimal feedback rule g = -A.!{3 than
under the nonfeedback rule g = O.
Sargent and Wallace observed that this argument in favor of feedback control
depends on the assumption that A and Pcan be taken to be independent of g.
However, suppose that (7.3) is generated by the structural model
(7.5)
and so use the currently prevailing gin [7.7] in forming its expectations, rather than the
old g that held during the estimation period. The public would presumably know g if
the monetary authority were to announce [it]. Failing that. the public might be able to
infer g from the observed behavior of the money supply and other variables. In any case,
on the assumption that the public knows what g the authority is using, ... A. ... comes
to depend on the authority's choice of g.
g changed and eventually agents figured this out, but that a modeling error was
made in specifying g to be a parameter in the first place. If it is believed that some
evidence will convince agents that g has changed, then this belief should be
incorporated in the model: g should be modeled as a variable rather thana para-
meter. One way to implement this alternative would be to replace (7.6) by a more
complex specification in which agents attach some prior probability to the event
of g changing, and then update their posterior distribution of g according to Bayes's
law. Such a respecification would allow the analyst to avoid the implication
that agents never revise their estimate of g if that specification is thought to be
implausible.
Sargent and Wallace, in contrast, effectively proposed to replace rational
expectations with stationary expectations: having "known" that g takes on a par-
ticular value in the estimation period, agents then come to "know" that g takes on
a different value as policy changes.
It was argued above that the treatment of policy regimes as parameters that vary
over time is inconsistent with rational expectations. To summarize, the point is
that if an analyst opts to model a regime as a parameter, rational expectations
implies that he or she must be willing to assume both that its value never changed
in the past and will never change in the future and further that the agents whose
behavior is being modeled know this to be the case. 5 Correspondingly, to model
regime change over time in a setting of rational expectations, one would assume
that regimes are the values taken on by a stochastic process the parameters of
which are known to agents. Lucas's reformulation of the Hall-Jorgenson model is
a good example of a successful analysis of policy regimes; another is Hood and
Garber (1983). Hamilton has several papers on stochastic regime switching, one
in this volume. It is seen that under rational expectations a policy regime is pro-
perly modeled as a state of a stochastic process that has a small number of states
and a low probability of state change at any date and not as a parameter that varies.
There is no qualitative distinction between policy regimes and policy variables,
and no harm would be done if the term regime were deleted.
How important is this point? Obviously the answer depends on the context. If
policy change is infrequent, little harm is done by representing policy regimes by
parameters. To insist on the distinction between rational expectations and station-
ary expectations in such settings would be excessively fastidious, since the two
would have virtually identical implications in practice and stationary expectations
is easier to implement analytically. But almost by definition most policy changes
are of a type that occurs frequently, and when policy change is frequent the regime
ON POLICY REGIMES 247
fonnulation is likely to lead to major error. Lucas, in fact, used exactly this ob-
servation to motivate his refonnulation of the Hall-Jorgenson analysis of the in-
vestment tax credit, as was observed above. In any case, the purpose of the foregoing
sections was not to argue that Lucas's preferred formulation of policy change can
never be used fruitfully but instead to question his contention that the alternative
fonnulation of the theory of economic policy is inherently flawed at a deep meth-
odological level.
Policy Ineffectiveness
Conclusion
Leo Rogin (1956) stressed that developments in economic theory can be under-
stood only in relation to the economic problems and policy questions that these
developments are intended to clarify. For example, Rogin observed, the extended
Ricardo-Malthus controversy on various points of pure theory can be rendered
intelligible only if it is situated relative to the differing positions the two took
on the overriding policy question that engaged both: should the Com Laws be
repealed? In contrast to Rogin, new classical macroeconomists contend that, if
anything, the opposite is true: controversy about ideological questions and policy
ON POLICY REGIMES 249
issues represents the ephemera of economics generally. not its deepest meaning
(for example, Lucas, 1980). The quality of economists' work as scientists, we are
told, is limited primarily by the scope and sophistication of the intellectual tools
at their disposal, and the enduring contribution of new classical macroeconomics
owes to its development and use of powerful analytical tools not available to
earlier macroeconomists.
The arguments of this paper support Rogin: policy questions, not technical
issues, are at the heart of new classical macroeconomics. The energy of new
classical macroeconomics springs from its rejection of the Keynesian mode of
policy conduct and policy advice, not from such technical innovations as rational
expectations, important as these are. The idea that technocrats in Washington can
guide the economy toward a noninflationary full-employment equilibrium by im-
plementing subtle adjustments in macroeconomic policy settings strikes new clas-
sical macroeconomists as antidemocratic, besides being implausible. It is difficult
not to sympathize: the image of academic economists jetting down the coast
to Washington to whisper in the ear of policy makers at congressional hearings
appears elitist even to those entirely free of Midwestern populist leanings.
The central project of new classical macroeconomics was to discredit dis-
cretionary macroeconomic policy and the associated mode of policy advice by
economists. Analytical developments in new classical macroeconomists are best
understood when situated in relation to this project. The Lucas critique and the
policy ineffectiveness proposition directly address policy questions; these there-
fore must be seen as the core propositions of new classical macroeconomics.
Rational expectations, not being directly related to policy questions, is secondary.
In fact, the theme of this paper is that the Lucas critique and policy ineffectiveness
proposition are derived under stationary expectations, not rational expectations.
The secondary role of rational expectations in new classical macroeconomics is
amply demonstrated by the fact that it was sacrificed in favor of stationary expec-
tations in arguing the Lucas critique and policy ineffectiveness proposition.
Macroeconomists were initially drawn to new classical macroeconomics be-
cause it appeared to provide a unified explanation for a series of perceived macro-
economic policy failures. Later the cost of embracing this explanation became
evident: new classical doctrine failed to equip macroeconomists with analytical
tools enabling them to think in a disciplined way about what policy response might
be appropriate to current macroeconomic conditions. Not only this; the instruction
of new classical macroeconomics was that the policy question framed this way is
inherently misposed. Macroeconomists simply did not buy this argument. Culture-
bound or not, most wanted to think about policy in relation to current economic
conditions and were led shortly to look elsewhere for guidance about how to do
so. The argument of this paper is that macroeconomists have displayed a sound
instinct in rejecting the policy analysis of new classical macroeconomics: nothing
250 MACROECONOMETRICS
Notes
1. The material in this paper was presented in Cooley, LeRoy, and Raymon (1984). See also Sims
(1982), (1986), Grossman (1984), and Cooley (1985).
2. For example: "I shall argue ... that simulations using [the major econometric) models can, in
principle, provide no useful information as to the actual consequences of alternative economic policies.
[This contention) ... will be based not on deviations between estimated and" 'true' structure prior to
a policy change but on deviations between the prior 'true' structure and the 'true' structure prevailing
afterwards" (Lucas, 1976, p. 105).
3. Another argument adduced in support of rules is that if policy rules are simple and if changes
in policy are well understood, there is reason to hope that agents' decision environments will be
rendered regular enough to be amenable to economic modeling. Complex policy rules and ambiguous
changes in these rules result, it is held, in situations of Knightian uncertainty in which it is impossible
for agents to act rationally and therefore difficull for economists to predict their actions. (Singell and
I, 1987, have questioned the attribution of this idea to Knight; whether or not Lucas's citation of Knight
is appropriate is irrelevant in the present context, however.) Here it is presumed that complex decision
rules for the government translate directly into complex decision environments for private agents. This
is assuming what needs to be demonstrated-namely, that there do not exist exogenous shocks that
governments can succeed in offsetting. In the contrary case a successful feedback control rule would
simplify agents' environment relative to that resulting from a simple control rule and would therefore
presumably render their behavior easier to predict, not more difficult.
4. I have simplified by deleting constants throughout.
5. In a setting of asyrnmetric information, the corresponding assumption is that the fact that the
parameter does not change over time is the common knowledge of the agents being modeled.
6. New classical macroeconomists, while insisting on the importance of explicitly modeled opti-
mization in characterizing the behavior of private agents, have not applied the same reasoning to
governments. This is a puzzling asymmetry. Surely the similarities between, say, the Federal Reserve
Board and the General Motors Corporation outweigh their differences.
References
Blanchard, Olivier, and Danny Quah. (1989). "The Dynamic Effects of Aggregate Demand
and Supply Disturbances." American Economic Review 79,621-636.
Cooley, Thomas F. (1985). "Individual Forecasting and Aggregate Outcomes: A Review
Essay." Journal of Monetary Economics 15, 255-266.
Cooley, Thomas F., Stephen F. LeRoy, and Neil Raymon. (1984). "Econometric Policy
Evaluation: Note." American Economic Review 74,467-470.
ON POLICY REGIMES 251
Flood, Robert P., and Peter M. Garber. (1983). "A Model of Stochastic Process Switching."
Econometrica 51, 537-551.
Friedman, Milton. (1948). "A Monetary and Fiscal Framework for Economic Stability."
American Economic Review 38, 245-264.
Friedman, Milton. (1957). A Theory ofthe Consumption Function. Princeton, NJ: Princeton
University Press.
Grossman, Herschel I. (1984). "Counterfactuals. Forecasts and Choice-Theoretic Model-
ling of Policy." Mimeo, Brown University.
Hall, Robert E. Jr., and Dale W. Jorgenson. (1967). "Tax Policy and Investment Behavior."
American Economic Review 57, 391-414.
Hoover, Kevin D. (1989). "The Logic of Causal Inference: Econometrics and the Condi-
tional Analysis of Causation." Mimeo, University of California, Davis.
Keynes, John Maynard. (1936). General Theory ofEmployment, Interest and Money. New
York: Harcourt, Brace & World.
leRoy, Stephen F. (1991). "Causal Orderings." Chapter 6 in this volume.
leRoy, Stephen F., and Larry J. Singell. (1987). "Knight on Risk and Uncertainty." Journal
of Political Economy 95, 394-406.
Lucas, Robert E. Jr. (1976/1981). "Econometric Policy Evaluation: A Critique." In Robert
E. Lucas, Jr., Studies in Business-Cycle Theory. Cambridge, MA: MIT Press.
Lucas, Robert E. Jr. (I 980a/l 98 I). "Methods and Problems in Business Cycle Theory." In
Robert E. Lucas (ed.). Studies in Business-Cycle Theory. Cambridge, MA: MIT Press.
Lucas, Robert E. Jr. (l980b/1981). "Rules, Discretion and the Role of the Economic Ad-
visor." In Robert E. Lucas, Jr., Studies in Business-Cycle Theory. Cambridge, MA: MIT
Press.
Mankiw, N. Gregory. (1990). "A Quick Refresher Course in Macroeconomics." Journal of
Economic Literature 28, 1645-1660.
Rogin, Leo. (1956). The Meaning and Validity of Economic Theory. New York: Harper.
Sargent, Thomas J. (1984). "Autoregressions, Expectations and Advice." American Eco-
nomic Review 74, 408-415.
Sargent, Thomas 1., and Neil Wallace. (1976/1981). "Rational Expectations and the Theory
of Economic Policy." In Robert E. Lucas, Jr. and Thomas Sargent (eds.), Rational
Expectations and Econometric Practice. Minneapolis: University of Minnesota Press.
Simon. Herbert A. (1953). "Causal Ordering and Identifiability." In William C. Hood and
Tjalling C. Koopmans (eds.), Studies in Econometric Method.
Sims, Christopher 1. (1982). "Policy Analysis with Econometric Models." Brookings Papers
on Economic Activity.
Sims, Christopher J. (1986). "Are Forecasting Models Usable for Policy Analysis?" Quar-
terly Review (Federal Reserve Bank of Minneapolis) 10,2-16.
Tinbergen, Jan. (1952). On the Theory of Economics Policy. Amsterdam: North Holland.
Commentary on Chapter 7
William Roberds
Introduction
The difference between my view of the Lucas critique and LeRoy's (1994) can
perhaps best be described as a sort of cultural gap between two generations of
economists. LeRoy sees the critique as a frontal attack on the validity of policy
analysis as traditionally practiced by economists and particularly on the theory of
economic policy advanced by Tinbergen and others. This was the view of the
critique that was prevalent when I entered graduate school in 1978, and this view
is still held by a large segment of the profession today.
However, the generation of economists trained since the publication of the
Lucas critique has been forced to view the critique in less epic terms. After all, the
publication of the critique did not spell the end of "life among the policy econ."
Papers, dissertations, and memos containing policy analyses continue to be written
by the truckload. Many of these analyses follow LeRoy's suggestion and ignore
the critique completely, others strictly hew to the path outlined by Lucas, and still
others fall between these two extremes. Yet it is safe to say that since 1976, the
interpretation of all such analyses has been tempered by the arguments advanced
by Lucas and by the counterarguments advanced in papers such as LeRoy's. While
it is still unresolved whether the critique represents a stepping stone on the way
to policy analysts' nirvana, no economist today can write as if the critique had
never existed.
In what follows, I will try to present my own rather prosaic interpretation of the
issues surrounding of the disagreement between the Lucas critique and the various
papers that criticize it, including LeRoy's. It is my opinion that the essential
disagreement can be quickly stated using an idea put forth in a 1971 paper by H.S.
Witsenhausen-that is, the notion of causality as it applies to information struc-
tures. I Witsenhausen' s idea of causality is also useful in explaining why neither
the Lucas view nor the critics' view amounts to a wholly satisfactory approach to
the problems of economic policy analysis.
253
254 MACROECONOMETIUCS
Above the vector y, describes the state of the economy and evolves according to
Here x, represents the decision variable(s) of the policy maker, (J, the decision
variable of the private sector, and E, a purely random shock. The decision variables
x, and (J, are chosen on the basis of the information available to the policymaker
and the private sector, respectively, available at time t. For analytical convenience,
partition the policymaker and the private-sector representative agent into a se-
quence of policymakers {p,} and the representative agent into a sequence of
representative agents {r,}. The time t policymaker p, chooses the value of x, at time
t, where x, must depend on the information available:
x, = G,(lf).
Similarly, the time t representative agent r, chooses (J, based on information
available:
(J, = H,(l ~).
The most natural definition of the information sets of the policymakers and the
representative agents will include the (field generated by the) history of the game
{(y" x" (J" E,)} up to time t. I assume that the policymaker will physically move
first in every period.
Most of the argument about the Lucas critique can be described as an argument
over the appropriate specifications of information for the policymaker (or se-
quence of policymakers) I~ or the information for the representative agents l~. In
general it is difficult to interpret these specifications of a game's infonnation
structure unless they satisfy a condition similar to Witsenhausen's (1971) causal-
ity property. Intuitively, the causality property requires "that for any play of the
game the actions of the agents can be ordered [so] that the information available
to an agent may depend on the decisions of agents acting earlier but cannot depend
COMMENTARY ON CHAPTER 7 255
upon the decisions of agents acting at concurrent or later times." In other words,
the causality property formalizes the idea that agents may not condition their
actions on things that have not yet happened. Mathematically, the causality prop-
erty expresses conditions that are sufficient for a given mathematical game to be
written down in extensive or "tree" form.
To illustrate the idea of causality, consider a simple variant of the policy game
outlined above. In this simple game there is no uncertainty and a finite horizon of
two periods. 3 To simplify things further, assume that the second period private
agent rz has only one possible decision and that the other players (PI' rl' pz) in the
game have only two possible decisions: call these H and L for high or low in-
flation. Then the set of all possible outcomes of the game is given by the eight
n
possible sequences for (PI, r l , pz)-that is, the set = {ilL, UH, LHL, HU,
LHH, HHL, HLH, HHH}. The set of all possible information about is ~ the n
power set of .Q; agents' information sets will be subsets of fJi. Under the most
natural specification of the policy game, the information set of the first-period
policy maker is Ii = {¢, .Q}, the information set of the first-period private agent
is 1[= {¢, D., {LU, UH, LHL, LHH}, {HLL, HHL, HLH, HHH}} , and the
information set of the second-period policy maker is If = Ii U {{UL, UH},
{LHL, LHH}, {HU, HLH}, {HHL, HHH}}. Note that if the game is played in the
order (PI' rl,pz), then IfC I~ C If, which for this example means that the causality
condition is satisfied for that particular order of play.
In the context of this example, my interpretation of LeRoy's argument (the fifth
section) is that the above specification of information sets is the natural and correct
one for policy analysis. This specification corresponds to our commonsense notion
of causality-that is, of time running in only one direction. More precisely, the
assumed temporal order of play corresponds to the nesting of information sets that
satisfies the causality restriction. However, Lucas evidently has a different speci-
fication in mind. In Lucas's formulation of the policy game, the policymakers PI
and pz would make their decisions before the private agent rl' Since If:::> Ir, how-
ever, the causality property would be not be satisfied by the ordering of players
(PI' Pz, r l )· On the other hand, if policy decisons are required to be time-invariant,
then the second policy maker's information set can be respecified as I~* == If, and
the causality condition will be satisfied for the order of play (PI' Pz, r,), since If
~ If* C Ir. Under this new specification of information sets, the causality con-
dition is not satisfied for the order of play preferred by LeRoy-that is, (PI' rl' pz).
debate over infonnation structures ("rules of play") of policy games, rather than
any property of equilibria such as subgame perfection or time-consistency. Second,
both Lucas's and LeRoy's preferred infonnation structures fonnally satisfy the
causality condition, although each satisfies causality through a different order of
play. Third, by Witsenhausen's (1971) Theorem 1, both specifications are for-
mally consistent with rational expectations. Witsenhausen's theorem states that if
the causality property is satisfied (for some order of play), then it is possible to
construct sequences of agents' decisions such that these are mutually consistent
and such that all agents' expectations are well defined.
The essence of LeRoy's criticism of Lucas's approach is that it is counter-
intuitive. In tenns of infonnation structure in the Lucas version of foregoing
example, LeRoy's criticism is manifested by the (1) the restriction of the second-
period policy-maker's infonnation set and (2) the changed order of play necessary
to satisfy the causality property. The infonnation structure preferred by LeRoy
satisfies causality without requiring restrictions on future policymakers' ability to
act on new infonnation and without requiring a counterintuitive order of play.
On the other hand, LeRoy is correct to suggest that one of Lucas's purposes in
fonnulating a policy game in this nonintuitive fashion was to argue the superiority
of rules over discretion. This point was made more forcefully by Kydland and
Prescott (1977). In the language of the preceding section, Kyd1and and Prescott
showed that some policy games played by Lucas's rules would lead to more
desirable outcomes than policy games played under the more intuitive specifica-
tion preferred by LeRoy.
There is also a certain amount of intuitive support for the Lucas-Kydland-
Prescott view. If one considers the real-world legal codes of most civilized soci-
eties, an important function of all such codes is to guarantee a degree of inertia in
most governmental policy decisions. In matters of policy, continuity for continu-
ity's sake evidently has a value that must be balanced against the current wisdom,
no matter how compelling that wisdom may be.
While the causality property is not a necessary condition for the existence of
rational expectations equilibria,4 it is a defensible contention that policy rules be
derived under some restriction that approximates the causality property. Optimal
rules derived for noncausal infonnation structures tend to have the following sort
of character. Consider a dynamic optimal taxation problem of the sort analyzed
by, for example, Sargent (1987, pp. 419-438). Given sufficient initial private
wealth, then the unconstrained solution to such a problem is always a one-time tax
tantamount to a confiscation of wealth sufficient to finance all future government
expenditures. Such counterintuitive results can be avoided, to some degree, by
restricting policies to be parametric-for example, by constraining tax policies to
be invariant in the capital taxation example. A restriction to time-invariant policy
rules that forces policy analysis into a causal framework can be understood as an
COMMENTARY ON CHAPTER 7 257
LeRoy and allied papers make a valid point when they, in effect, argue that
Lucas's methodology violates our commonsense notions of temporal order and
causality. Yet Lucas and his supporters have a point when they argue that in a
policy environment with rational agents, there is something to be gained by con-
sidering policy rules, which satisfy causality only under conditions that are im-
plausible. An ideal theory of economic policy would be able to capture potential
welfare gains associated with policy credibility, without elevating policy makers
beyond the everyday constraints imposed by our physical notions of time and
space.
Since the publication of the critique, several threads in the macroeconomic
literature have attempted to formulate such a theory of economic policy, with
mixed results. The most prominent such branch of the literature, typified by papers
such as Barro and Gordon (1983), Rogoff (1987), and Chari, Kehoe, and Prescott
(1989), has presented more refined versions of macropolicy games similar to the
one described above. For example, Barro and Gordon (1983) describe a macropolicy
repeated game with an information structure that satisfies the causality restriction
in the natural way preferred by LeRoy. An equilibrium of this repeated game
corresponds to a time-inconsistent equilibrium of a one-shot version of this game.
The Barro-Gordon result suggests that certain time-inconsistent policies that
are equilibria of games with counterintuitive information structures could be
rationalized as equilibria of games with more intuitively appealing information
structures.
Rogoff (1987) points out a major disadvantage associated with this approach,
however. This drawback is familiar to even casual students of game theory as the
problem of "too many equilibria." In most cases it is possible to write down many
reasonable variants of a given macropolicy game, each with many reasonable
equilibria. As Rogoff (1987, p. 165) notes, "it would seem premature to focus
attention only on the most favorable equilibria."
Chari, Kehoe, and Prescott (1989) show that the Barro-Gordon type results can
be derived for truly dynamic policy games involving play between a policy maker
and a representative private agent, whose actions mirror potential conflicts of in-
terests among private agents. Chari, Kehoe, and Prescott are careful to restrict the
information structures of the games they consider to be causal along the lines pre-
ferred by LeRoy (in their words, strategies are history-contingent.) Once again they
show that certain time-inconsistent equilibria can be at least closely approximated
258 MACROECONOMETIUCS
Such stories are not damning evidence against this new class for models, but at a
minimum they point to the need for further refinement.
Conclusion
In the face of the theoretical ambiguity outlined above, how should policy analysis
proceed? leRoy (fifth section) suggests that correct approach is to just ignore the
Lucas critique, as long as one's econometric model is capable of capturing some
low-probability, low-frequency shifts in the nature of the stochastic processes that
the being modeled. He argues that distinction between policy regimes and policy
variables is an artificial one.
My conjecture is that leRoy's prescription is for now probably the correct one
for many of the marginal, short-term problems of policy analysis that are routinely
faced by many macroeconomists and particularly for the large segment of the
profession who work for governmental agencies. Economists in these positions
are often called upon to construct multiple what-if scenarios on short notice. If the
policy changes under consideration are marginal and the planning horizons are
short-term, then there is no reason to expect that Lucas-type policy analyses with
structural models would give more accurate answers.
However, many economists continue to see the old-fashioned, purely econo-
metric approach to policy analysis as insufficient for predicting the longer-term
consequences of major changes in policy. In such cases, the absolute accuracy of
econometric projection is not so much the issue as an ability to interpret and
defend the results. Along this dimension a Lucas-type policy analysis is likely
to be at an advantage, any logical inconsistencies notwithstanding, over a more
empirical approach.
My impression is that this ugly but pragmatic viewpoint is shared by the ma-
jority of economists trained since 1976, across the ideological spectrum. Despite
leRoy's contention that macroeconomists have rejected the policy analysis of the
new classical macroeconomics, one sees more and more of the Lucas-type analy-
sis in applied settings. For example, ten years ago policy discussions at Federal
Reserve System meetings usually consisted of arguments over money-demand
regressions, but now the same discussions often center on Lucas-type interven-
tions in applied general equilibrium models. This trend is likely to continue given
the diminishing costs of calculating and estimating rational expectations equilibria.
Certainly leRoy and the other critics of the Lucas critique have advanced our
profession's understanding of the critique and have done the profession a service
by pointing out some of the unappealing features of Lucas's approach to policy
analysis. In the view of a growing subset of the profession, however, these prob-
lems are not so great as to outweigh the insights gained by Lucas's approach. Until
260 MACROECONOMETRICS
Acknowledgments
Thanks are due to Jerry Dwyer, Eric Leeper, and Chuck Whiteman for comments
on earlier drafts, with the usual disclaimer. I am especially grateful to Mike Stutzer
for making me aware of Witsenhausen's work. The opinions expressed are my
own and not those of the Federal Reserve Bank of Atlanta or the Federal Reserve
System.
Notes
I. It is unfortunate that Witsenhausen's concept shares the name causality with the econometric
idea of causality. The two concepts are related but distinct.
2. For the purposes of the discussion below, it is not critical that the policy maker seeks to
maximize any objective. The use of a representative agent for the private sector, though customary
in the literature, glosses over some important issues. See Chari, Kehoe, and Prescott (1989) for a
discussion.
3. As shown in Witsenhausen (1971), the applicability of the causality condition is quite general
and is not restricted to such simple examples. Above I consider the simplest possible case due to space
limitations.
4. Witsenhausen's (1971) Theorem 2 shows that if there are more that two players (or stages) in
a game, solvability and measurability of a given solution of an information structure does not imply
that the structure satisfies causality.
5. On this possibility see Sims (1988).
References
Barro, Robert J., and David B. Gordon. (1983). "Rules, Discretion and Reputation in a
Model of Monetary Policy." Journal of Monetary Economics 12 (July), 101-121.
Chari, V.v., Patrick J. Kehoe, and Edward C. Prescott. (1989). "Time Consistency and
Policy." In Robert J. Barro (ed.), Modem Business Cycle Theory. Cambridge, MA.:
Harvard University Press.
Kydland, Finn E., and Edward C. Prescott. (1977). "Rules Rather Than Discretion: The
Inconsistency of Optimal Plans." Journal of Political Economy 85 (June), 473-491.
Leroy, Stephen F. (1994). "On Policy Regimes." Chapter 7 in this volume.
Rogoff, Kenneth. (1987). "Reputational Constraints on Monetary Policy." In Karl Brunner
and Allan H. Meltzer. (eds.), Carnegie-Rochester Conference Series on Public Policy,
Vol. 26, Bubbles and Other Essays. Amsterdam: North-Holland.
COMMENTARY ON CHAPTER 7 261
Introduction
263
264 MACROECONOMETItlCS
This section considers the general economic structure postulated by Lucas, which
a simple expectations model illustrates. The Lucas critique is empirically refutable
as well as confirmable, this section also examines tests of the Lucas critique and
then considers what constitutes empirical evidence for and against the Lucas
critique.
Whether the Lucas critique applies for a specific economic relationship is thus an
empirical issue. Specifically, two related properties of (J are of concern: its con-
stancy (or lack thereof) as I\. changes and its invariance to (or dependence on) 1\..1
As suggested by Gordon (1976, pp. 48-49) and Neft~i and Sargent (1978) and
developed by Hendry (1988), Favero and Hendry (1992), and Engle and Hendry
(1993), these properties provide two approaches to testing the Lucas critique:
266 MACROECONOMETRICS
Several remarks are in order concerning tests of the Lucas critique. Specifically,
clarification of what does and does not constitute empirical evidence on the Lucas
critique is critical, since many of the articles in the literature survey below present
evidence that is uninformative on the empirical applicability of the Lucas critique.
First, the empirical observations motivating Lucas's critique could arise from
more mundane causes, such as dynamic misspecification, omitted variables, and
incorrect functional form. Because those causes cannot be precluded a priori, non-
constancy of conditional models and vector autoregressions is uninformative about
the Lucas critique. Such nonconstancy could arise from the Lucas critique applying.
THE LUCAS CRITIQUE IN PRACTICE 267
Equally, nonconstancy could arise from, for example, an omitted variable bias,
where that bias depended on a correlation with a policy variable and so would
change when the associated policy rule changed.
Second, estimation of an expectations-based model is not evidence in itself for
or against the Lucas critique, even if the implied cross-equation restrictions hold.
The parameterization of a conditional-marginal representation could be such that
those restrictions hold as well.
Third, and relatedly, modeling of expectations does not necessarily resolve the
Lucas critique, even when the Lucas critique applies. If expectations are improp-
erly modeled, policy analysis with the corresponding misspecification of (8.1) and
(8.2) may well result in inaccurate and biased simulations. McCallum (1989,
p. 230), a strong proponent of rational expectations models, emphasizes this point:
"[U]sing models based on explicit optimization analysis of individual agents'
choices ... can provide no guarantee of success; explicit optimization analysis
will not help if the agent's objective function or constraints are misspecified. The
true message of the Lucas critique is simply that the analyst must be as careful as
possible to avoid the use of equations that will tend, for plausible reasons, to shift
with policy changes." To establish the empirical applicability of the Lucas critique,
a researcher would require demonstrating that alternative explanations for the
failure of a conditional model did not apply. Equally, the proposed expectations-
based model would need to explain the data adequately-that is, be empirically
congruent in the sense of Hendry and Richard (1982). For system-based expec-
tational models, empirical congruence is required for (8.2) as well as (8.1): in par-
ticular, the changes in policy regime must be modeled. This contrasts with much
of the empirical literature on rational expectations, where models are estimated
assuming that A is constant even though the Lucas critique assumes that A is not.
Fourth, Sims's (1987) modification of Lucas's framework does not appear
empirically relevant for a wide range of relationships examined through the litera-
ture searches in the third through fifth sections of this chapter or for the empirical
example in the fifth section. Noting a potential logical inconsistency in Lucas's
framework, Sims proposes allowing A to be time-varying, with its distribution
depending upon some fixed metaparameter. That implies that X, and y, are ergodic,
coming from a distribution with constant parameters. Such a model is at odds with
the observed nonconstancy of many empirical equations; see Judd and Scadding
(1982) on money demand alone. Also, the potential logical inconsistency may
vanish in the presence of infonnation costs, as implied by Cochrane's (1989)
analysis and the fifth section below.
Fifth, if the Lucas critique is empirically refuted by tests 1 or 2 above, the
refutation is generic in the sense that a whole class of expectations-based models
is inconsistent with the data evidence. This class includes typical rational expec-
tations models and also a wide variety of expectations-based models where the
268 MACROECONOMETItlCS
To assess the empirical importance of the Lucas critique, we searched the Social
Science Citation Index (SSCI) for articles citing Lucas (1976) and examined the
THE LUCAS CRITIQUE IN PRACTICE 269
empirical evidence in those articles. This section summarizes how the literature on
the Lucas critique was assembled and what the shortcomings of that assemblage
are.
This subsection discusses the construction of a database on the literature about the
Lucas critique. The interest is in obtaining as complete a list as possible of articles
discussing the Lucas critique and in categorizing that literature according to vari-
ous characteristics. The Social Science Citation Index (SSCI) allows searches for
occurrences of a particular article as a reference in another article. So, we searched
the SSCI for all articles published during 1976-1990 that included Lucas (1976)
in their list of references. 3 The resulting citations were obtained in a computer-
readable form and formatted as a database in Borland's (1992) Paradox for Win-
dows. For each citing article, the initial database included the author, date of
publication, title, journal, journal volume and number, and page numbers.
In total, 590 articles cite Lucas (1976). We were able to obtain copies of over
96 percent of those articles. Of those obtained, fifty-six are treated as "bad cites"
because they do not cite Lucas (1976) in the text of the article, do not cite Lucas
(1976) in the bibliography, are not in English, or are not journal articles. Table 8.1
details the breakdown. This filtering results in 513 articles available with "good
cites" to Lucas (1976).
The articles were read and analyzed, and additional information about the
articles was added to the database, as described in Table 8.2. A complete listing
of the information on the articles appears in Ericsson and Irons (l995b, app.).
Briefly, the articles were categorized according to the nature of the article, the
context of the article's citation to Lucas (1976), the author's view on whether or
not the article's evidence confirmed the Lucas critique, and the sorts of evidence
presented on the Lucas cirtique. These categorizations require some additional
explanation:
I. The nature of an article is either theoretical, empirical, mixed (both
theoretical and empirical), or "other" (book review, survey, or interview).
Many articles develop a theory model and illustrate it with an empirical
application. Depending on the originality of the theory and the substan-
tiveness of the empirical application, such an article could be considered
theoretical, mixed, or empirical.
2. The context of the cite to Lucas (1976) is either tangential, postulated, or
substantial. Numerous articles mention the Lucas critique only in passing
and are thus considered "tangential." In these articles, Lucas (1976) typi-
cally is referenced only in a footnote, in a sentence about previous work
270 MACROECONOMETRICS
Table 8.1. Availability, status. and nature of the articles from the SSGI search on
Lucas (1976)
Articles available:
Good cites:
Empirical 143
Mixed 147
Theoretical 135
Other 88
Subtotal 513
Bad cites:
Lucas (1976) not cited in the text 24
Lucas (1976) not listed in the bibliography 8
Article not in English 19
Book 5
Subtotal 56
Subtotal 569
Articles not available 21
All articles in the SSeI search 590
3. The author's views on the applicability of the Lucas critique are catego-
rized as yes, maybe, no, assumed, warning, or not applicable ("NA"). We
have tried to preserve the author's own views, as presented in the paper,
regardless of the validity or type of test employed. If the author believes
that the Lucas critique is empirically relevant, given the evidence pre-
sented (whatever that evidence might be), then the article is categorized
as "yes." The "yes" and "assumed" categories are sometimes only subtly
different. If the author presents any type of evidence, including past cites,
the paper is categorized as "yes." If no evidence is presented, yet it is
clear that the author thinks the critique is relevant, the paper is classified
as "assumed." Some authors take no view on the applicability of
the Lucas critique ("NA") or merely cite Lucas (1976) as a "warning" to
modelers.
272 MACROECONOMETIUCS
With the augmented database, tabulation by one or more types of classification can
shed light on the empirical evidence on the Lucas critique.4 Before doing so in
the fourth section, various shortcomings of the methodology and the data are
examined.
The methodology for gathering and constructing the data is closely related to the
data themselves, so shortcomings of the methodology and of the data are consid-
ered together. Shortcomings include possible sample selection bias in the SSCI
search and in our use of the SSCI search, and fuzziness in categorization.
The SSCI search itself entails sample selection of several sorts. First, the issue
of the Lucas critique may appear in an article without Lucas ever being mentioned.
Because the Lucas critique was widely discussed well prior to Lucas's own for-
mulation (as in Haavelmo, 1944), relevant articles prior to 1976 are excluded.
Second, the Lucas critique may be mentioned explicitly in the text, but Lucas
(1976) may not appear in the references. In many circles, "Lucas critique" is a
household word, and citation to the paper itself may be missing. Third, the SSCI
excludes some publications citing Lucas (1976), such as books and articles
appearing in books.
For the first two sources of sample selection, the sins of omission are no doubt
great although the direction of bias is unclear. Correcting for those omissions is
simply beyond the scope of this paper and the current technology for citation
searches. The third source is considered immediately below in conjunction with
sample selection in our use of the SSCI.
As implied by Table 8.1, our use of the SSCI entails additional sample selec-
tion, which may bias our findings. Most but not all of the articles from the SSCI
search were obtained: 569 out of 590. The twenty-one articles "not available"
either were not in the libraries accessible to us or were not available in a reason-
able amount of time. However, all articles citing Lucas (1976) and appearing in
THE LUCAS CRITIQUE IN PRACTICE 273
the 130 most cited economic journals are "available."s The articles "not available"
are in journals very infrequently cited by the economics profession, such as Potato
Research and Transportation Research Part B (Methodological). Fifty-six of the
available articles are "bad" because, for example, they list Lucas (1976) in their
references but do not cite Lucas in the body of the paper. The remaining 513
articles are considered "good." Tables 8A.l, 8A.2, and 8A.3 (in the Appendix)
respectively list the numbers of good, bad, and unavailable articles by journal. In
short, the percentage of articles available is very high, and the articles obtained
reflect the core of published economic research.
Categorization is an issue with some articles. In particular, both the nature of
the article and the context of the cite to Lucas (1976) can be difficult to decide-
for example, as in classifying an article as tangential or postulated. Some fuzziness
may exist at the edges of the associated types, but overall this does not appear to
be a substantial problem.
This section summarizes the statistical characteristics of the citation database and
examines it for empirical evidence in favor of the Lucas critique. Surprisingly
little evidence exists that substantiates the Lucas critique in practice, so this sec-
tion also considers reasons for that lack of evidence.
The Evidence
70 I
60 -
50
'"c0
"§
40
'u
'-
0
....
Il,)
.0 30
e::s
Z 20
10
OL---l.----'----'-~-'----'---'-----'-----'----'-----'-----'---'--'-----L-J
Table 8.3 tabulates the joint distribution of cites across the nature of the article,
the context of the cite, and the view of the author. While this ignores the time
structure in Figures 8.2 through 8.4, Table 8.3 provides the basis for winnowing
the citations. Of the 513 good citations, 327 are tangential and 98 are postulated.
Of the remaining eighty-eight substantial citations, twenty-nine are purely theo-
retical and five are surveys or the like, leaving fifty-four empirical or mixed
articles that cite Lucas (1976) in a substantive context. Of those, only forty-three
find any evidence for the Lucas critique: less than 10 percent of the total citations.
Table 8.4 lists the types of empirical evidence on the Lucas critique provided
by those forty-three citations. Each category of evidence and its informativeness
on the Lucas critique is now considered.
Eighteen articles show the nonconstancy of a conditional model and five that
of a vector autoregression (VAR) or reduced form. As such, this evidence on
nonconstancy is uninformative about the empirical validity of the Lucas critique.
For instance, a vector autoregression may be nonconstant even while a factoriza-
tion of the associated joint density of the data may have a constant-parameter
conditional equation. The VAR of x, and y, from (8.4) and (8.5) could be nonconstant
due to changes in A even if (J is constant and is invariant to A. Likewise, the
THE LUCAS CRITIQUE IN PRACTICE 275
'"~
tJ
15 15
10
"t
:in'n ~. 8
5
,
I 'L
"
o '. .
1976 1978 1980 1982 1984 1986 1988 1990 1976 1978 1980 1982 1984 1986 1988 1990
: : r, -----:.=.::..::.::=.-----_1 Theoretical
20
Other
, 15 .
10 10
0lL.J.-,,",-,-,,",-,-,,",-,-,,",-,..L.l...L.l...L.l...L.l...L..L..L..L..L..L..L..L..L..LJ-LJ
1976 1978 1980 1982 1984 1986 1988 1990 1978 1980 1982 1984 1986 1988 1990
Figure 8.2. Yearly number of citations to Lucas (1976), by nature of the article
:r ..
Substantial Postulated
·1
:~ i
.
:f
20 .
o l=..LJ..JLl...<:::I..LcLLLLLLl..LL.LLLJ..JLJ...!..LLJ...LlJ
1976 1978 1980 1982 1984 1986 1988 1990 1978 1980 1982 1984 1986 1988 1990
Tangential
Figure 8.3. Yearly number of citations to Lucas (1976), by context of the citation
276 MACROECONOMETIUCS
Yes Maybe
20,----------------, 20,---------'--------...,
15 15
10 ... 10
5 5
0~.!.'_'_.L..1_'_'_'U_''_LJ._"_'_~''_'__'_''
-i
...........L..1_l-WL..L_U OU-L.!.'_'_J...LL-L.'-"-'-'-'_"_'_...L..L"-L..!.'_'_J....J....J....ULLJ....L.LJ
1976 1978 1980 1982 1984 1986 1988 1990 1976 1978 1980 1982 1984 1986 1988 1990
No Assumed
20,---------------, 20,---------------,
15 .......•........ 15 ...........•..
10 10
5 - . 5
1976 1978 1980 1982 1984 1986 1988 1990 1976 1978 1980 1982 1984 1986 1988 1990
Warning NA
20 r-----------~ 20,----------------,
15 15
10 10
5 5 rr .
OU-LJ...L~'_'_''_LJ._"_'_...L..L...L..L''_'_.!.'_'_J...LJ....J....L...I....l...J-U fln~ 11 II
OU-1.-'..L-'..LJ.....l...L..l-L.J...J-L..I'_LJ....J....l....L..L..LL.L.L.L.L.LL.J.J
1976 1978 1980 1982 1984 1986 1988 1990 1976 1978 1980 1982 1984 1986 1988 1990
Figure 8.4. Yearly number of citations to Lucas (1976), by view of the author on
the empirical validity of the Lucas critique
THE LUCAS CRITIQUE IN PRACTICE 277
Table 8.3. Author's view on the empirical applicability of the Lucas critique: By
nature of the article and context of the cite
Substantial:
Yes 19 10 13 42
Maybe 9 5 5 19
No 4 6 I II
Assumed g I 9
Warning I 2
NA 3 2 5
Subtotal 32 22 29 5 88
Postulated:
Yes 3 8 7 2 20
Maybe 5 4 2 6 I7
No I I I 3
Assumed 7 17 12 4 40
Warning 4 2 2 8
NA 2 7 10
Subtotal 22 32 22 22 98
Tangential:
Yes 7 10 1 18
Maybe 8 7 6 5 26
No 5 5 1 11
Assumed 6 II 9 26
Warning 30 24 9 19 82
NA 33 36 58 37 164
Subtotal 89 93 84 61 327
Total 143 147 135 88 513
Table 8.4. Empirical evidence presented for the Lucas critique: By nature of the
article and type of evidence
Empirical Mixed
Nonconstancy:
Conditional model 8 3 4 3 18
Reduced fonn or VAR I 2 I I 5
Rational expectations model I I 2
Noninvariance:
Conditional model 0
Reduced fonn or VAR 2 2
Rational expectations model 0
Prediction or simulation 2
Survey data 5 7
Cross-equation restrictions 2 2 5
Constant rational expectations model
and nonconstant alternative model 2
Total 19 9 10 5 43
Note: Articles included above make a substantial cite to Lucas (1976). are either empirical or
mixed, and have an author's view of yes or maybe.
Peek and Wilcox (1987) construct an empirically constant model for the determi-
nation of U.S. interest rates over 1952-1984. Their model contains expected prices,
which are calculated from the Livingston survey data on inflation expectations,
and allows for the effects of changing monetary regimes. While constancy of the
model using the expectations data is of economic interest, it does not confirm or
disprove the Lucas critique. Rather, its constancy shows the econometric value of
the survey data. The measured expectations could be adaptive rather than forward
looking or could be derived from a data-based (rather than model-based) predictor.
Articles modeling the survey data itself provide no information on the Lucas
critique, since inter alia the importance of the measured expectations in agents'
decisions is not shown.
Five articles corroborate the Lucas critique by finding statistically acceptable
cross-equation restrictions in expectational models. As noted in the second section
above, those restrictions could be satisfied by a conditional-marginal representa-
tion as well.
Each of the final two articles finds an empirically constant expectations model
and (on the same data) an empirically nonconstant alternative model. This evid-
ence on the Lucas critique appears more fruitful, so these articles are considered
in greater detail.
One of the articles (McNelis and Neft~i, 1982) examines two output supply
models for the United States over 1954-1978, one being a conditional equation
subsequently published in Fair (1984, p. 132) and the other being Sargent's (1976)
"supply-side alternative" involving expectations in prices. The models are estim-
ated with the Kalman filter, allowing for time-varying coefficients, similar to the
recursive estimation procedure employed in the fifth section. The estimated para-
meters of Sargent's equation appear "somewhat more stable over time" (p. 305)
than those in Fair's equation. While suggestive, these results can only be taken as
tentative. First, both models show some apparent nonconstancy, although no formal
tests of constancy are calculated, partly from the (then) scarcity of good econometric
techniques for doing so. Much better tests are now available (see Hoffman and
Pagan, 1989; Ghysels and Hall, 1990; Hansen, 1992; Dufour and Ghysels, 1996).
Second, the coefficient on unanticipated price changes in Sargent's model is sta-
tistically insignificant for virtually all subsamples and becomes (barely) signifi-
cant at the 95 percent level at the very end of the sample. The unanticipated price
change is the only expectational variable in Sargent's model, so the role of expec-
tations in obtaining constancy in Sargent's model is questionable. Finally, Sargent's
model would need to be subjected to the standard battery of diagnostic tests over
an extended dataset to establish its empirical viability.
The other article (LaHaye, 1985) accounts for expectations of currency reform
in Cagan's model of money demand, using data from hyperinflationary periods in
Germany, Poland, and Greece. For all three countries, Cagan's model is empirically
280 MACROECONOMETIUCS
constant if expectations of currency refonn are incorporated; and for Gennany and
Greece, the model is nonconstant if those expectations are excluded from it. With
expectations of currency refonn included, LaHaye's algorithm for estimating
Cagan's model in effect treats the date of refonn as known (pp. 548-550). Such
knowledge mayor may not have been available to agents prior to the refonn. In
any case, a constant expectational model does not necessarily preclude a condi-
tional model that is constant also (see Hendry, 1988; Ericsson and Hendry, 1989).
In summary, the vast majority of articles citing Lucas (1976) are not concerned
with testing the Lucas critique per se. Those that are provide scant evidence for
its empirical basis.
In order to understand the lack of empirical evidence for the Lucas critique, it is
useful to summarize how authors have dealt with the critique. Three approaches
dominate.
In the first, the Lucas critique is ignored, or claims are made that it is of little
significance to the problem at hand. For example, many authors argue that, despite
the theoretical complications of the Lucas critique, their policy or counterfactual
simulations are valid because the policy experiments fall within the range of his-
torical experience and thus do not constitute a regime change. Other authors warn
that the results of their policy simulations are not strictly interpretable because of
the Lucas critique, but then proceed to calculate and discuss these results, ignoring
the Lucas critique. Still others find empirically constant conditional models and
thus conclude that the Lucas critique must not be a problem, albeit without check-
ing for changes in the marginal process or determining what those changes are.
In the second approach, the author acknowledges that the Lucas critique may
apply. It is then circumvented by avoiding periods of structural change or by in-
cluding dummies, or the empirical results are used for another purpose. For exam-
ple, several authors interpret empirical nonconstancy in an equation as indicating
that a structural shift occurred in the economic relation being modeled. Here, the
Lucas critique is cited in defining a regime change, rather than for using a regime
change to investigate the importance of the Lucas critique. Many papers set out
to determine whether a regime shift in variable x occurred in year t without
examining the effect of such a regime shift on any particular model. Quite com-
monly, the sample is split according to a priori reasoning (for example, in half)
or in light of an ex post regime shift, and estimation results over the subsamples
are then compared. The Lucas critique is used to explain why the estimated
parameters may have changed. As discussed in the second section above, the
Lucas critique is only one explanation for parameter instability: many other types
THE LUCAS CRITIQUE IN PRACTICE 281
This section examines empirical evidence against the Lucas critique by consider-
ing evidence on super exogeneity. It illustrates the procedures for testing super
exogeneity with a particular empirical model, Hendry and Ericsson's (199Ia, eq.
(10» U.K. money demand function. Then, it examines the evidence on super
exogeneity generally, developing and analyzing a citation databas~ for super
exogeneity.
282 MACROECONO~CS
The Lucas critique implies that conditional models cannot have super exogenous
variables, leading to the two testable implications discussed in the second section
above: the constancy of parameters in conditional models and the invariance of
those parameters to changes in the marginal process. To clarify the practical
application of the associated tests, this section analyzes super exogeneity in Hendry
and Ericsson's'(1991a) conditional model of annual U.K. money demand. This
subsection shows that that model is empirically constant. The following subsec-
tions demonstrate the nonconstancy of the marginal processes, thus showing super
exogeneity via test 1; extend the marginal models for test 2; calculate test 2,
finding super exogeneity for the conditional model and demonstrating the high
empirical power of the invariance test on misspecifications of the conditional
model; and consider some implications of the conditional model and associated
tests.
Using Friedman and Schwartz's (1982) annual data, Hendry and Ericsson
(1991a) obtain the following conditional error correction model:
-
~(m - p), = 0.45~(m - P),_I - O.lO~2(m - P)'-2 - O.60~,
[0.06] [0.04] [0.04]
where
T = 93 [1878-1970], R 2 =0.87, f1 = 1.424%.
The variables are per capita nominal money M2 (M), real net national product (Z),
its deflator (P), short- and long-term nominal interest rates (RS and RL), and
dummies for the two world wars (D 1 and DII ), with logs indicated by lower case.
Details of the data appear in Friedman and Schwartz (1982) and Hendry and
Ericsson (1991a). The residualU is the error correction m - P - z + 0.309 + 7.ooRS,
and it enters (8.6) nonlinearly. White's (1980) heteroscedasticity consistent stand-
ard errors are in brackets, ~ is the difference operator, T is the number of obser-
vations, R 2 is the squared multiple correlation coefficient, f1 is the estimated equation
standard error, and a circumflex indicates least squares estimation.
A
Hendry and Ericsson (1991a) show that (8.6) passes a battery of diagnostic
tests and that it is constant over subsamples prior to and after World War I.
Recursive estimation over the entire sample was not undertaken because of the
dummy (D 1 + D II ), in (8.6). Constancy of the conditional model is central to the
THE LUCAS CRITIQUE IN PRACTICE 283
.836
.83
......... ........... .............. ................... ... ,., .
.... .....
, .
• 824
.818
.812
-.818
-.824 ...
/,,, ,. . . . . .
-. 83 L.....~1-9-8..L8~~19~1....l8~~1~9-28
.........~1~9 3..L8~~19~4:-'8~~··_·~...:.~':'~~......~~1~9...
6...L8~~1-:"97
....8........~
Figure 8.5. One-step residuals (-) from the money demand equation (8.6), with
o ± 20', (...)
first super exogeneity test, so (8.6) is estimated recursively over the entire sample,
taking the full-sample estimate of the coefficient on (DJ + DJI), as known.6 Figure
8.5 plots the one-step residuals and the corresponding equation standard errors-
that is, {y, - {J:x,} and {O ± 2a,} in a common notation. The equation standard
error a varies little over time. Figure 8.6 records the "break-point" Chow (1960)
statistics for the sequence {1893-1970, 1894-1970, ... , 1969-1970, 1970}. None
of the Chow statistics are significant at their one-off 5 percent critical values.
Likewise, Hansen's (1992) test against coefficient nonconstancy with an unknown
break point is insignificant at the 5 percent level for each coefficient, for the
equation error variance, and for the coefficients and equation error variance jointly.
Figures 8.7 and 8.8 plot the recursive estimates and standard error bands for co-
efficients on two central variables, current inflation [Ap,] and the error correction
term [(U'_I - 0.2)u~_I]. Both coefficients are empirically constant over time; the
other coefficients in (8.6) show similar stability. The marked narrowing of the
standard error bands in the 1890s and just after World War I reflects the high
information content of the data during those periods. Results in the following
284 MACROECONOMETRlCS
J..J.
J. ........................................................••.......•..•..........................................................
.9
.8
.7
r
.6
.5
.4
.3
.2
.J.
9
J.999 J.9J.9 J.929 J.939 J.949 J.959 J.969 J.979
Figure 8.6. Break-point Chow statistics (-) for the money demand equation (8.6),
normalized by their one-off 5 percent critical values (...)
Inflation and the two interest rates enter (8.6) contemporaneously, so the
nonconstancy of their marginal processes is of interest for testing super exogeneity.
Figure 8.9 plots the annual inflation rate and the long-tenn interest rate, while
Figure 8.10 plots the two nominal interest rates. Several features of the data are
striking. First, inflation deviates wildly from RL during and after World War I,
albeit in opposite directions for the two subsamples. Second, the ex post real long
rate is generally positive but does not appear constant. Third, the short- and long-
tenn nominal interest rates tend to move together, excepting the period 1932-
1952. Thus, inflation was modeled as a univariate marginal process, whereas RS
and RL were modeled both as univariate processes and as a bivariate marginal
process, the latter permitting cointegration between RS and RL. This subsection
THE LUCAS CRITIQUE IN PRACTICE 285
.2
'.
_ --'-'<;._-_ _ - - _.__ __._-_ _.__ _ __ _._ _ _.
"
-.2
-.4 ......
~ ,.,. -,...... ",
., \ : _- ,
-.6
t····,,·············..······..··..····.. . .
-.8
//'\"
-J.
" ~ .. -
' .""'.:"'/""
\/
J.9""
Figure 8.7. Equation (8.6): recursive estimates (-) of the coefficient of inflation
IIp,, with ±2 estimated standard errors (...)
presents Ericsson and Irons's (1995a) marginal process for inflation and develops
the univariate marginal models for the interest rates. The next subsection develops
the bivariate marginal model for the interest rates.
Starting with a fifth-order autoregressive model ofPr' Ericsson and Irons (1995a)
obtain the following generalized random walk:
/"'0.
tlpr = 0.65tlPr_1 + 0.0081, (8.7)
[0.18] [0.0036]
where
T = 93 [1878-1970], R 2 = 0.43. fJ = 4.239%, Inn F(4, 87) = 1.22.
Inn F(·, .) is the standard F statistic for testing that the reduction from the more
general (fifth-order autoregressive) model is valid. Under the null hypothesis of
valid reduction, Inn F(·, .) is asymptotically distributed as an F statistic with
degrees of freedom as given (albeit ignoring the presence of a unit root in p,). The
reduction to (8.7) appears statistically acceptable. Still, (8.7) is highly nonconstant.
286 MACROECONOMETRlCS
-8
-111
-12
-14'-"-~1~9-88
.........~1-9~1 .....
8~~19~2..J.II~-19~3.....8~~1~94-J8........~1-9-58
..........~1-9-6..L.8~-1-9:"-:?-:-8~
Figure 8.8. Equation (8.6): recursive estimates (-) of the coefficient of the error
correction term (0,-1 - O.2)t1r_l' with ±2 estimated standard errors (...)
Figures 8.11 and 8.12 plot the recursively estimated equation standard errors and
the break-point Chow statistics.
A similar picture develops for the interest rates. Starting from univariate fifth-
order autoregressive models for RS and RL, the following simple models were
obtained:
----
tJ.RS, = -0.37tJ.RSr_2 + 0.00078,
[0.13] [0.00091]
(8.8)
where
.2:1
.2
.J.:I
-.8:1
-.J.
-.J.:I
-. 2 L...-.~"""""~~-,-~.-....""""'~"""""""""~""""'~-"-""""""'~'-'-~~......J..~-,-,....L-~"'-'--'
J.888 J.898 J.988 J.918 1928 1938 1948 J.958 J.968 1978
Figure 8.9. Annual rate of inflation !!.pI (-) and the long-term interest rate RL, (...)
where
Figures 8.13 through 8.16 plot the recursively estimated equation standard errors
and the break-point Chow statistics for each of (8.8) and (8.9). Equation (8.8) may
be constant, but (8.9) clearly is not. For the latter, Hansen's (1992) joint instability
test is 1.92 with four degrees of freedom, rejecting at the I percent level. Using
test I from the second section above, super exogeneity of prices and interest rates
in (8.6) follows from the constancy of the conditional model (8.6) and the
nonconstancy of the marginal models (8.7) and (8.9).
To perfonn test 2 from the second section above, empirically more constant,
better-fitting marginal models are developed for !1p, RS, and RL. Equations (8.7)
through (8.9) are extended to include dummies, which proxy for shifts in A. over
288 MACROECONOMETRICS
.J.
o
J.88& J.89& J.9&& J.9J.& J.92& J.930 J.94& J.95& J.96& 197&
Figure 8.10. The short-term and long-term interest rates, RS, (-) and RL,
(...)
time and which thus are used in the next subsection to test the invariance of the
coefficients in (8.6). The major shifts are of several forms, so a notation for
dummy variables is convenient. lab denotes an impulse dummy for the year 19ab
(that is, being +1 in 19ab and zero otherwise). Sabcd denotes a step dummy begin-
ning in the year 19ab and ending in the year 19cd (that is, being + I for 19ab-19cd
and zero otherwise). Some dummies also interact with the lagged dependent vari-
able or with a time trend (trend). The dummies aim to capture the inflationary and
deflationary periods during and after World War I (SIS20' S2I23) and the large
deviations between RS and RL for part of the interwar period and during and
following World War II (S33S1' S3945' 132 , 133 , 152 ). Additionally, dummies are in-
cluded for the remainders of the respective samples (S2470, Smo).
Ericsson and Irons (1995a) model inflation as the first-order autoregressive
process (8.7) augmented by three dummies (S1520' S2I23, S2470), which enter directly,
interactively with inflation, and interactively with the time trend. Aiming to maxi-
mize the power of the invariance test, Ericsson and Irons simplify that augmented
equation to obtain the following:
THE LUCAS CRITIQUE IN PRACTICE 289
.12
............. _- ...
. .
••• ••••••• Oh ••
. 08
.04
-.04 .
-.08
.................. .................
-.12
-.16
-.2
-.24
Figure 8.11. One-step residuals (-) from equation (8.7) for top,. with 0 ± 20',
(...)
where
T = 93 [1878-1970], R2 = 0.83, it = 2.388%, Inn F(5, 81) = 0.97.
OLS estimated standard errors (in parentheses) are reported for coefficients on
dummy variables, since the dummies often have implausibly small estimated
heteroscedasticity consistent standard errors. Equation (8.10) is more constant
than (8.7) and has an estimated equation standard error it virtually half that of
(8.7), indicating the statistical and numerical importance of the additional vari-
ables in (8.10).
The interest rates RS and RL are modeled as a bivariate fifth-order VAR in light
of their apparent cointegration in Figure 8.10. That VAR is estimated with a
290 MACROECONOMETRICS
4.4
~.6
Figure 8.12. Break-point Chow statistics (-) for equation (8.7) of lip,. normalized
by their one-off 5 percent critical values (...)
constant only (System A), with a constant and Sml (System B), and with a con-
stant and all dummies associated with the interest rates (that is, Sml' S3945, S5270,
Smo . trend, 132 , 133 , 152 ) (System C). The dummies enter the systems unrestrictedly
and are allowed to (but not forced to) enter the cointegrating vector, if one exists.
While System A does not appear to be cointegrated, both Systems B and Care:
see Table 8.5. Inclusion of Sml appears critical in capturing the temporary devia-
tion between RS and RL around and immediately prior to and after World War n. 7
Using Johansen's (1988, 1991) and Johansen and Juselius's (1990) maximum
likelihood procedure, the null hypothesis of no cointegration is rejected at the 1
percent level in favor of one cointegrating vector for Systems B and C. The
estimated cointegrating vector is approximately (I, -0.7), and the estimated error
correction feedback coefficients in the equations for RS and RL are approximately
-0.7 and -0.03, respectively. From Johansen's (l992a, 1992b) test, RL appears
weakly exogenous for the cointegrating vector (that is, - 0.03 '" 0), whereas
RS clearly is not weakly exogenous. Long-run unit homogeneity is marginally
rejected in System C, although not when tested jointly with the weak exogeneity
of RL.
THE LUCAS CRITIQUE IN PRACTICE 291
.82
""" ............................
...............'
.815
..: . " ,' .
..." , .
-.815 ................
. .
-.82
-.825
Figure 8.13. One-step residuals (-) from equation (8.8) for RS,. with 0 ± 2c1,
(... )
With these data features in mind, System C was simplified to the following pair
of equations, which are estimated by maximum likelihood:
........---
tJ.RS, = 0.53tJ.RL'_1 - 1.l9tJ.RL,_2 - 0.613(RS - RL),_1 - 0.0006
[0.24] [0.36] [0.093] [0.0011]
--
- 0.0158S33sl , + 0.0047S394S• - 0.025/32" (8.11)
(0.0026) (0.0028) (0.008)
where
T = 93 [1878-1970], d"RS = 0.743%, d"RL = 0.238%, corr(uRS' URJ = 0.63,
Inn X2(22) = 30.5 [0.11].
292 MACROECONOMETIUCS
1.2
.8
.6
.4
.2
Figure 8.14. Break-point Chow statistics (-) for equation (8.8) of RSt , normalized
by their one-off 5 percent critical values (...)
The statistic corr(uRS' URJ is the empirical correlation between the residuals from
the two equations, and Inn X 2(22) is the likelihood ratio test statistic against the
unrestricted System C. Cointegration is readily apparent from the large and highly
significant feedback coefficient on (RS - RL)t_1 in (8.1l), which is statistically
close to the corresponding value of -0.81 from the unrestricted VAR in System
C. If the error correction term is added to (8.12), it is statistically insignificant and
numerically small, in line with RL being weakly exogenous. Lagged changes in
the long-term interest rate affect both short- and long-term rates, whereas lags in
the short-term rate matter only through the error correction term, which appears
in only (8.11). Several of the dummy variables are statistically and numerically
highly significant, reflecting their importance in capturing some of the major
movements in the series. While (8.11) and (8.12) are more constant and better
fitting than (8.8) and (8.9), equations (8.11) and (8.12) still have some residual
autocorrelation and heteroscedasticity. An even more general model of interest
rates might capture those features of the data: for the present purposes, (8.11) and
(8.12) suffice.s
THE LUCAS CRITIQUE IN PRACTICE 293
.liU
.998
.996
.................................
• 994
..........
• 992
-.992
-.994 .............................
-.996
-.998
Figure 8.15. One-step residuals (-) from equation (8.9) for RL" with 0 ± 2<1/
(...)
Using equations (8.10) through (8.12), this subsection constructs the second type
of super exogeneity test for (8.6) and provides evidence on the empirical power
of that test. The column under "(8.6)" in Table 8.6 lists four variants of that test
statistic for super exogeneity. The first three variants test for the significance
in (8.6) of the right-hand side variables from the marginal models (8.10), (8.11),
and (8.12), respectively, with redundant variables excluded from the test. The final
statistic (using "all" marginal models) tests for the joint significance of all three
marginal models' right-hand side variables. None of the four statistics are signifi-
cant, even at the 20 percent significance level: asymptotic p-values appear in
square brackets. The parameters in (8.6) appear invariant to the changes in the
inflation and interest rate processes, so inflation and interest rates are super
exogenous in (8.6), precluding a role for model-based expectations of those vari-
ables in U.K. money demand.
Power may be an issue for both super exogeneity tests 1 and 2, and their finite
sample power can be assessed directly. Hendry and Ericsson (1991a) show that the
294 MACROECONOMET~CS
14
191i19
Figure 8.16. Break-point Chow statistics (-) for equation (8.9) of RLt • normalized
by their one-off 5 percent critical values (...)
* Significant at 5 percent.
** Significant at I percent.
Notes:
I. The systems are fifth-order VARs in RS and RL over 1878-1970. All systems include a constant
term. In addition, System B includes S3"'; and System C includes SJJS1' S394" SS2700 Smo' trend, In,
133 • and Is>, Asymptotic p-values appear in square brackets.
2. The statistics A••" and Au"" are Johansen's maximal eigenvalue and trace eigenvalue statistics
for testing cointegration. The null hypothesis is in terms of the cointegration rank r and, e.g., rejection
of r = 0 is evidence in favor of at least one cointegrating vector. The statistics A:". and A.~"." are the
same as A...,. and A."""" but with a degrees-of-freedom adjustment. Critical values are taken from
Osterwald-Lenum (1992, table I).
3. The weak exogeneity test statistics are evaluated under the assumption that r = I and so are
asymptotically distributed as X2( I) if weak exogeneity of the specified variable for the cointegrating
vector is valid. The joint test statistic is asymptotically distributed as X2(2) if RL is weakly exogenous
and long-run unit homogeneity holds.
296 MACROECONOMETIUCS
Table 8.6. Invariance test statistics for super exogeneity in equation (8.6)
Misspecification in lag
!J.p F(5,79) 0.78 11.9** 0.88 0.78 13.4**
[0.56] [0.000] [0.50] [0.57] [0.000]
RS F(6,78) 0.67 1.49 2.29* 2.81 * 5.39**
[0.67] [0.19] [0.044] [0.016] [0.000]
RL F(6,78) 1.12 0.30 1.73 2.83* 2.60*
[0.36] [0.93] [0.13) [0.015) [0.024)
All F(14,70) 1.12 5.22** 2.02* 1.72 8.18**
[0.35] [0.000] [0.028] [0.07] [0.000]
(t 1.424% 2.710% 1.541% 1.500% 3.349%
Misspecification by omission
fjp F(5,80) 9.99** 0.78 0.80 11.0**
[0.000] [0.56] [0.56] [0.000]
RS F(6,79) 1.51 2.41* 1.44 4.88**
[0.19] [0.034] [0.21] [0.000]
RL F(6,79) 0.31 1.79 1.95 2.01
[0.93] [0.11 ) [0.08] [0.07]
All F(14,71) 4.91 ** 2.09* 1.10 7.25**
[0.000) [0.022) [0.38) [0.000)
& 2.700% 1.537% 1.508% 3.314%
* Significant at 5 percent.
** Significant at 1 percent.
Notes:
I. The marginal models for 6p, RS, and RL are given by equations (8.10), (8.11), and (8.12). The
right-hand side variables in those equations provide the basis for the various invariance tests.
2. Listed degrees of freedom in the F statistic are for (8.6) and for misspecifications involving 6rs
or 6 2 rl. Degrees of freedom for other misspecifications may differ slightly because 6p appears in (8.6)
at the first lag as well as current-dated. Asymptotic p-values appear in square brackets.
instance, misspecification in the lag of the short-term interest rate implies replac-
ing 6.rs, in (8.6) by 6.rs'_I' The invariance test statistic using the right-hand side
variables from (8.11) is F(6, 78) = 2.29 with a p-value of 4.4 percent (that is, the
second pair of values in the fifth column of Table 8.6). Much stronger rejection
results when the role of inflation in (8.6) is misspecified. The rejections in Table
THE LUCAS CRITIQUE IN PRACTICE 297
8.6 reflect the power of the test and the high infonnation content in the data. In
summary, inflation and interest rates are super exogenous for the parameters in the
money demand equation (8.6), and the power of the super exogeneity tests appears
high.
Some Implications
A few comments may help clarify the implications of these results. First, both
types of super exogeneity tests indicate that inflation and interest rates are super
exogenous in the money demand model (8.6). Instrumental variables estimation of
(8.6) by Hendry and Ericsson (l991a, p. 30) obtains virtually identical and empiri-
cally constant coefficient estimates, thus also supporting valid conditioning. Even
with super exogeneity, data-based predictors are allowable, so error correction
models such as (8.6) can have a forward-looking interpretation (see Hendry
and Ericsson, 1991b). Specifically, current and lagged inflation enter (8.6) as
-O.3(~PI + ~2PI) (approximately), where ~PI + ~2PI is a natural predictor of next
period's inflation. Flemming (1976) proposes similar such models for fonning
expectations about inflation; Campos and Ericsson (1988) and Kamin and Ericsson
(1993) generalize the class of predictors and implement them in conditional models
of Venezuelan consumers' expenditure and Argentine broad money demand re-
spectively. Super exogeneity is feasible with such a predictor because the pre-
dictor does not require estimating A.. That is, the predictor is data-based, not
model-based.
Second, conditional models may imply trivial losses in utility relative to an
expectational solution. Cochrane (1989) provides numerical comparisons for con-
sumers' expenditure, as do Akerlof (1979), Akerlof and Milbourne (1980), and
Hendry (1995, p. 582) for money demand. Data-based predictors may be only
slightly less accurate than model-based ones; and the infonnation costs to agents
may be high for achieving a model-based predictor that betters a data-based one,
particularly in the presence of frequent regime changes to a complicated policy
reaction function. A conditional error correction model thus may parallel an
optimal decision rule by agents facing such infonnation costs. More generally, Ss-
type models provide the economic underpinnings for error correction models such
as (8.6): see Akerlof (1979), Akerlof and Milbourne (1980), Milbourne (1983),
and Smith (1986) for money demand in particular.
Third, policy can and (in general) does affect agent behavior when super
exogeneity holds. Policy does so through the variables entering the conditional
model, albeit not through the parameters of that model. Government policy might
well affect inflation and interest rates, and so the demand for money. However,
under super exogeneity, the precise mechanism that the government adopts for
298 MACROECONOMETIUCS
such a policy does not affect agent behavior, except insofar as the mechanism
affects actual outcomes.
Fourth, Lucas (1988, p. 162) argues that the long-run money demand function
should be stable, even if its short-run behavior is not. The results above show that
both long- and short-run aspects of the estimated money demand function are
empirically constant and are invariant to the policy changes that occurred. 9 Its
empirical constancy is surprising, given Judd and Scadding's (1982) survey of
empirical money demand equations for the United States. The empirical constancy
of Kamin and Ericsson's (1993) conditional money demand equation for Argen-
tina is even more surprising, since its estimation period spans a sample with in-
flation rates ranging from under 1 percent to nearly 200 percent per month. Further,
the Lucas critique is of potential issue for even long-run parameters in conditional
money demand models, noting the numerous cites in Laidler (1993) to articles
modeling money demand with expectations for each of its right-hand side vari-
ables. "Policy" then should also be interpreted in Lucas's (1976, p. 21) broad
sense because government agencies are likely to be not the only forces determin-
ing income, prices, and interest rates.
Fifth, while the bivariate model of interest rates suffices for the expository
purposes of this empirical illustration, a marginal analysis is feasible for interest
rates, prices, and income jointly, as in Hendry and Ericsson (1986). More gener-
ally, analysis of the system for money, prices, income, interest rates, and (possi-
bly) other variables may offer additional insights for policy, even when super
exogeneity holds for the conditional money demand equation. On postwar quar-
terly U.K. data, Hendry and Mizon (1993) and Hendry and Doornik (1994) find
an additional cointegrating vector linking inflation with output relative to trend;
and the short-run dynamics of the marginal equations are important for under-
standing the implied behavior of money. Similarly, Juselius' s (1993) system analy-
sis of Danish money data helps clarify the interactions between variables, beyond
that obtainable from a single equation approach. By working with subsystems, our
analysis of U.K. data bridges Hendry and Ericsson's (1991a) single equation
results and an analysis of a closed system.
Table 8.7. Availability, status, and nature of the articles from the ssel search on
the exogeneity papers
Articles available:
Good cites:
Empirical 74
Mixed 44
Theoretical 39
Other 24
Subtotal 181
Bad cites:
Exogeneity papers not cited in the text 9
Exogeneity papers not listed in the bibliography 14
Article not in English 3
Book 4
Subtotal 30
Subtotal 21l
Articles not available 5
All articles in the SSeI search 216
Table 8.8. Author's view on the empirical presence of super exogeneity: By na-
ture of the article
Yes 18 7 25
Maybe 4 2 6
No 4 1 5
Weak exogeneity 4 4 8
NA 44 30 39 24 137
Total 74 44 39 24 181
Table 8.9. Empirical evidence presented for and against super exogeneity: By
type of test and area of application
Consumption 2 3
Expenditure 1 1
Housing 3 3
Money demand 4 10 7 5 15 2
Prices 1 1
Public finance 2
Trade
Wages
Total II 13 3 9 3 6 0 25 5
Note: The column "cond." indicates that the conditional model is empirically constant but that
evidence on the marginal process was not presented. Articles where the evidence was mixed or where
only weak exogeneity was tested are not included in the table. The two right-hand side columns list
the corresponding numbers of articles cited in a given row, which need not be the sums across the row,
noting that some articles report more than one of the tests.
Conclusions
This paper shows that the Lucas critique is a possibility theorem, not an existence
theorem. An extensive search of the literature reveals virtually no evidence dem-
onstrating the empirical applicability of the Lucas critique. That search clarifies
what sort of evidence would constitute support for the Lucas critique: much of the
evidence cited in the literature is uninformative because it could arise from more
mundane sorts of model misspecification. A case study of U.K. money demand
shows how super exogeneity tests can be used to test for and refute the Lucas
critique in practice. An additional literature search finds considerable evidence
refuting the Lucas critique for specific macroeconomic equations across a number
of countries. Model-based expectations may be important for some (other) areas
of the economy: that is an empirical issue, and one on which there is currently
scant evidence. Equally, rejection of super exogeneity for a particular conditional
model does not preclude another conditional model (explaining the same data)
from having super exogenous variables.
Refutation of the Lucas critique does not mean that agents are backward look-
ing rather than forward looking. Rather, refutation may imply that agents' forma-
tion of expectations is simply not model-based. The conditional model of U.K.
302 MACROECONOMETIUCS
Tables 8A.I, 8A.2, and 8A.3, respectively, list the numbers of good, bad, and
unavailable articles, by journal.
Table 8A.1. Number of articles in ssel searches for Lucas (1976) and the
exogeneity papers: Article is available and cite is "good"
Lucas Exogeneity Journal
8 Economic Inquiry
13 10 Economic Journal
13 3 Economic Modelling
9 Economic Record
3 2 Economica
I Economics and Philosophy
4 4 Economics Letters
11 7 European Economic Review
2 Explorations in Economic History
I Giornale degli Economisti e Annali di Economia
2 History of Political Economy
I Housing Finance Review
I Interfaces
6 4 International Economic Review
7 2 International Journal of Forecasting
I International Journal of Industrial Organization
9 2 International Monetary Fund Staff Papers
2 International Regional Science Review
I International Statistical Review
I Journal of Accounting Research
5 4 Journal of Applied Econometrics
I I Journal of Banking and Finance
I Journal of Business
4 3 Journal of Business and Economic Statistics
I Journal of Comparative Economics
I Journal of Contemporary Business
2 I Journal of Development Economics
I Journal of Development Studies
7 18 Journal of Econometrics
13 Journal of Economic Dynamics and Control
I Journal of Economic Education
Journal of Economic Issues
8 Journal of Economic Literature
4 Journal of Economic Perspectives
I Journal of Economic Studies
2 Journal of Economic Theory
I Journal of Economics and Business
I Journal of Economics (Zeitschrift fUr Nationalokonomie)
I Journal of Environmental Economics and Management
3 Journal of Finance
3 Journal of Forecasting
304 MACROECONOMETRICS
8 Weltwirtschaftliches Archiv
1 World Development
World Politics
Zeitschrift flir National6konomie (Journal of Economics)
513 181 Total
Table 8A.2. Number of articles in ssel searches for Lucas (1976) and the
exogeneity papers: Article is available and cite is "bad"
Lucas Exogeneity Journal
Table 8A.3. Number of articles in ssel searches for Lucas (1976) and the
exogeneity papers: Article is not available
Acknowledgments
Notes
I. Constancy need not imply invariance or invariance constancy. A parameter may be time-
varying yet still be invariant to policy interventions. Equally. a parameter may Jack invariance. yet be
constant because policy has remained unchanged over the sample.
2. Loosely speaking, x, is super exogenous for 8 in (8.1) if the parameters of interest can be
retrieved from 8 alone and if 8 does not depend on A or on the range of A. See Engle, Hendry, and
Richard (1983) for a precise definition of super exogeneity.
3. Searching the SSCI is not entirely straightforward, especially for Lucas (1976). The article
whose citation is to be searched is indexed by author, year, and journal, and there is little consensus
on the proper citation of Lucas (1976) or the proper way to index the cite. Sample indexes include
"Phillips Curve Labor" with 117 entries; "VI. P19, Carnegie-Rochester C ... " with sixty-three;
"J Monetary Ec S" with twenty-one; and dozens more with various permutations, (mis)spellings.
hyphenations, and abbreviations.
308 MACROECONOMETRICS
4. Some additional infonnation on the articles was gathered but is left for future analysis. For each
article, the SSCI provides the author's affiliation; and we established the pages on which Lucas (1976)
is cited, the article's primary and secondary JEL numbers, whether or not Lucas (1976) motivates the
article, whether or not Lucas (1976) is cited in the context of policy analysis, and whether or not the
article is about policy.
5. The frequency of journal citation is taken from Laband and Piette's (1994, Table A2) ranking
based on impact-adjusted citations per character for I990 citations to articles published in 1985-1989.
Their other rankings make little difference to the distribution between available and unavailable arti-
cles for citations to Lucas (1976) occurring in frequently cited journals.
6. Treating that coefficient as known may bias the Chow statistic towards rejection because an
extra degree of freedom is assumed. However, this bias did not matter in practice.
7. Other studies also include transients and short-run variables in their cointegration analyses.
Johansen and Juselius (1992) include oil price inflation in their analysis of purchasing power parity and
uncovered interest rate parity; Hendry and Mizon (1993) and Hendry and Doornik (1994) include
various dummies in their models of postwar U.K. money, prices, output, and interest rates; and Juselius
(1993) includes both transient dummies and stationary variables in her analysis of Danish money
demand. Kremers, Ericsson, and Dolado (1992) show analytically and empirically that accounting for
short-run or temporary fluctuations can dramatically improve the power of cointegration tests.
8. Univariate and bivariate models for the interest rates are qualitatively similar if logs of the
interest rates are modeled rather than their levels. Levels appear to deliver better-specified equations
than logs, so the levels equations are presented.
9. Hendry and Ericsson (I99la) show that the short-run elasticity of RS may have changed with
the introduction of Competition and Credit Control in 1971. That change does not obviate the empirical
constancy of the model over a near century of data, nor does it provide evidence for the Lucas critique.
Rather, the nonconstancy appears to arise from mismeasurement of the opportunity cost of holding
money (see Hendry and Ericsson, 1991a, p. 32). Lubrano, Pierse, and Richard (1986) and Steel and
Richard (1991) find comparable results in extensive Bayesian analyses of a similar U.K. aggregate
(M3); and Hendry and Ericsson's (I99lb) model ofMI demand is constant over a period spanning the
introduction of Competition and Credit Control.
10. Citations to unpublished versions of the papers are also included, noting that Engle, Hendry,
and Richard (1983) and Engle and Hendry (1993) were particularly long in gestation.
II. Fourteen articles in the SSCI search do not include any of the exogeneity papers in their
bibliographies, although they do cite other papers by Engle, Hendry, or Richard. Their inclusion in the
SSCI search reflects our inability in using the SSCI to specify fully a citation of interest when the
citation appeared as a discussion paper in a year different from its year of publication.
References
Akerlof, G.A. (1979). "Irving Fisher on His Head: The Consequences of Constant
Threshold-target Monitoring of Money Holdings." Quarterly Journal of Economics
93(2), 169-187.
Akerlof, GA, and R.D. Milbourne. (1980). "Irving Fisher on His Head II: The Conse-
quences of the Timing of Payments for the Demand for Money." Quarterly Journal of
Economics 95(1), 145-157.
Aldrich, J. (1989). "Autonomy." Oxford Economic Papers 41(1), 15-34.
THE LUCAS CRITIQUE IN PRACTICE 309
Baba, Y., D.F. Hendry, and R.M. Starr. (1992). "The Demand for Ml in the U.S.A., 1960-
1988." Review of Economic Studies 59(1), 25-61.
Blinder, A.S. (1988). "The Fall and Rise of Keynesian Economics." Economic Record 64,
187, 278-294.
Borland. (1992). Paradox for Windows Version 1.0: User's Guide. Scotts Valley, CA:
Borland International.
Borland. (1993). Quattro Profor Windows Version 5.0: User's Guide. Scotts Valley, CA:
Borland International.
Brodin, PA, and R. Nymoen. (1992). "Wealth Effects and Exogeneity: The Norwegian
Consumption Function 1966(1)-1989(4)." Oxford Bulletin of Economics and Statistics
54(3),431-454.
Campos, I., and N.R. Ericsson. (1988). "Econometric Modeling of Consumers' Expendi-
ture in Venezuela." International Finance Discussion Paper No. 325, Board of Gover-
nors of the Federal Reserve System, Washington, DC, Iune.
Chong, Y.Y., and D.F. Hendry. (1986). "Econometric Evaluation ofLinear Macroeconomic
Models." Review of Economic Studies 53(4), 671-690. Reprinted in C.W.I. Granger
(ed.), (1990), Modelling Economic Series: Readings in Econometric Methodology (ch.
17, pp. 384-410). Oxford: Oxford University Press.
Chow, G.c. (1960). "Tests of Equality Between Sets of Coefficients in Two Linear Regres-
sions." Econometrica 28(3), 591-605.
Cochrane, I.H. (1989). "The Sensitivity of Tests of the Intertemporal Allocation of
Consumption to Near-rational Alternatives." American Economic Review 79(3),319-
337.
Cuthbertson, K. (1991). 'The Encompassing Implications of Feedforward versus Feed-
back Mechanisms: A Reply to Hendry." Oxford Economic Papers 43(2),344-350.
Doornik, I.A., and D.F. Hendry. (1994). PcGive Professional 8.0: An Interactive Econo-
metric Modelling System. London: International Thomson.
Dufour, I.-M., and E. Ghysels (eds.). (1996). Journal of Econometrics. Special issue on
"Recent Developments in the Econometrics of Structural Change." In press.
Engle, R.F., and D.F. Hendry. (1993). "Testing Super Exogeneity and Invariance in Regres-
sion Models." Journal of Econometrics 56(112), 119-139.
Engle, R.F., D.F. Hendry, and I.-F. Richard. (1983). "Exogeneity." Econometrica 51(2),
277-304.
Ericsson, N.R. (ed.). (l992a). Journal of Policy Modeling, 14(3) and 14(4). Special issues
on "Cointegration, Exogeneity, and Policy Analysis."
Ericsson, N.R. (1992b). "Parameter Constancy, Mean Square Forecast Errors, and Meas-
uring Forecast Performance: An Exposition, Extensions, and Illustration." Journal of
Policy Modeling 14(4),465-495.
Ericsson, N.R., and D.F. Hendry. (1989). "Encompassing and Rational Expectations: How
Sequential Corroboration Can Imply Refutation." International Finance Discussion Paper
No. 354, Board of Governors of the Federal Reserve System, Washington, DC, Iune.
Ericsson, N.R., and I.S. Irons (eds.). (1994). Testing Exogeneity. Oxford: Oxford Univer-
sity Press.
Ericsson, N.R., and I.S. Irons. (1995a). "Book Review of Applied Econometric Techniques
310 MACROECONOMETRICS
the United Kingdom and the United States." European Economic Review 35(4), 833-
881.
Hendry, D.F., and G.E. Mizon. (1993). "Evaluating Dynamic Econometric Models by
Encompassing the VAR." In P.C.B. Phillips (ed.), Models, Methods, and Applications
ofEconometrics: Essays in Honor ofA.R. Bergstrom (ch. 18, pp. 272-300). Cambridge,
MA: Basil Blackwell.
Hendry, D.F., and J.-F. Richard. (1982). "On the Fonnulation of Empirical Models in
Dynamic Econometrics." Journal of Econometrics 20(1), 3-33. Reprinted in C.W,}.
Granger (ed.), (1990), Modelling Economic Series: Readings in Econometric Method-
ology (ch. 14, pp. 304-334). Oxford: Oxford University Press.
Hoffman, D., and A.R Pagan. (1989). "Post-sample Prediction Tests for Generalized Method
of Moments Estimators." Oxford Bulletin of Economics and Statistics 51 (3), 333-343.
Johansen, S. (1988). "Statistical Analysis of Cointegration Vectors." Journal of Economic
Dynamics and Control 12(2/3),231-254.
Johansen, S. (1991). "Estimation and Hypothesis Testing of Cointegration Vectors in
Gaussian Vector Autoregressive Models." Econometrica 59(6), 1551-1580.
Johansen, S. (I 992a). "Cointegration in Partial Systems and the Efficiency of Single-
equation Analysis." Journal of Econometrics 52(3), 389-402.
Johansen, S. (l992b). ''Testing Weak Exogeneity and the Order of Cointegration in UK
Money Demand Data." Journal of Policy Modeling 14(3),313-334.
Johansen, S., and K. Juselius. (1990). "Maximum Likelihood Estimation and Inference on
Cointegration-With Applications to the Demand for Money." Oxford Bulletin of Eco-
nomics and Statistics 52(2), 169-210.
Johansen, S., and K. Juselius. (1992). "Testing Structural Hypotheses in a Multivariate
Cointegration Analysis of the PPP and the UIP for UK." Journal of Econometrics
53(1/2/3), 211-244.
Judd, J.P., and J.L. Scadding. (1982). "The Search for a Stable Money Demand Function: A
Survey of the Post-1973 Literature." Journal of Economic Literature 20(3), 993-1023.
Juselius, K. (1993). "VAR Modelling and Haavelmo's Probability Approach to Macro-
economic Modelling." Empirical Economics 18(4),595-622.
Kamin, S.B., and N.R. Ericsson. (1993). "Dollarization in Argentina." International Finance
Discussion Paper No. 460, Board of Governors of the Federal Reserve System, Wash-
ington, DC, November.
Kremers, lJ.M., N.R Ericsson, and lJ. Dolado. (1992). "The Power of Cointegration
Tests." Oxford Bulletin of Economics and Statistics 54(3), 325-348.
Laband, D.N., and M.J. Piene. (1994). "The Relative Impacts of Economics Journals:
1970-1990." Journal of Economic Literature 32(2), 640-666.
LaHaye, L. (1985). "Inflation and Currency Refonn." Journal ofPolitical Economy 93(3),
537-560.
Laidler, D.E.W. (1993). The Demandfor Money: Theories. Evidence. and Problems (4th
ed.). New York: HarperCollins.
Lubrano, M., RG. Pierse, and l-F. Richard. (1986). "Stability of a UK Money Demand
Equation: A Bayesian Approach to Testing Exogeneity." Review of Economic Studies
53(4), 603-634.
312 MACROECONOMETRICS
Lucas, Jr., R.E. (1976). "Econometric Policy Evaluation: A Critique." In K. Brunner and
A.H. Meltzer (eds.), The Phillips Curve and Labor Markets, Carnegie-Rochester Con-
ference Series on Public Policy, Vol. I, Journal ofMonetary Economics, supplementary
issue, 19-46.
Lucas, Jr., R.E. (1988). "Money Demand in the United States: A Quantitative Review." In
K. Brunner and B.T. McCallum (eds.), Money, Cycles, and Exchange Rates: Essays in
Honor of Allan H. Meltzer, Carnegie-Rochester Conference Series on Public Policy,
Vol. 29, Journal of Monetary Economics, supplementary issue, 137-168.
McCallum, B.T. (1989). Monetary Economics: Theory and Policy. New York: Macmillan.
McNelis, P.D., and S.N. Neft~i. (1982). "Policy-dependent Parameters in the Presence of
Optimal Learning: An Application of Kalman Filtering to the Fair and Sargent Supply-
side Equations." Review of Economics and Statistics 64(2), 296-306.
Milbourne, R. (1983). "Optimal Money Holding Under Uncertainty." International Eco-
nomic Review 24(3), 685-698.
Neft~i, S.N., and TJ. Sargent. (1978). "A Little Bit of Evidence on the Natural Rate
Hypothesis from the U.S." Journal of Monetary Economics 4(2),315-319.
Osterwald-Lenum, M. (1992). "A Note with Quantiles of the Asymptotic Distribution of
the Maximum Likelihood Cointegration Rank Test Statistics." Oxford Bulletin of Eco-
nomics and Statistics 54(3), 461-472.
Peek, 1., and J.A. Wilcox. (1987). "Monetary Policy Regimes and the Reduced Form for
Interest Rates." Journal of Money, Credit, and Banking 19(3),273-291.
Sargent, TJ. (1976). "A Classical Macroeconometric Model for the United States." Journal
of Political Economy 84(2), 207-237.
Sims, C.A. (1987). "A Rational Expectations Framework for Short-run Policy Analysis."
In W.A. Barnett and KJ. Singleton (eds.), New Approaches to Monetary Economics
(ch. 14, pp. 293-308). Cambridge: Cambridge University Press.
Smith, G.W. (1986). "A Dynamic Baumol-Tobin Model of Money Demand." Review of
Economic Studies 53(3), 465-469.
Steel, M.FJ., and J.-F. Richard. (1991). "Bayesian Multivariate Exogeneity Analysis: An
Application to a UK Money Demand Equation." Journal ofEconometrics 49( 112), 239-
274.
White, H. (1980). "A Heteroskedasticity-consistent Covariance Matrix Estimator and a
Direct Test for Heteroskedasticity." Econometrica 48(4),817-838.
Commentary on Chapter 8
Eric M. Leeper
Introduction
313
314 MACROECONOMETRICS
characterize the class of policy interventions that economic theory predicts will
shift private behavior.
Viewing the Lucas critique as a theorem about economic behavior, and its
implication for the stability of econometric models as a potential application of the
theorem, leads inevitably to the issue of identification. Thus, the central theme of
this comment is summarized by the question: In their empirical example of money
demand, have Ericsson and Irons identified private and policy behavior that sup-
ports their conclusion that they can "refute the Lucas critique"? Although Ericsson
and Irons liberally use the phrase "refute the Lucas critique," no amount of em-
pirical evidence can disprove the logical proposition. More modestly, and more
accurately, the authors have found that a particular empirical relationship re-
mained stable in a particular data sample even though the stochastic processes for
the contemporaneous explanatory variables displayed instability.
This comment proposes some interpretations of what the authors found and dis-
cusses the usefulness of the authors' procedures for policy analysis. Before turning
to these interpretations, I will comment briefly on the literature searches that
Ericsson and Irons conduct.
Over half of the paper reports results from searches of the Social Science Citation
Index. Ericsson and Irons search for references to Lucas (1976) and to super
exogeneity tests. They focus initially on the fifty-four citations to Lucas that are
empirical in nature. Of those, forty-three claim to find evidence supporting the sort
of behavior that Lucas emphasizes. The authors then critically evaluate the forty-
three articles according to their notions of what constitutes "valid" empirical
evidence. 3 The authors' evaluation includes phrases like: "a vector autoregression
may be nonconstant even while a factorization of the associated joint density of
the data may have a constant-parameter conditional equation," "the nonconstancy
of a particular conditional model does not preclude some other conditional model
from being constant," "measured expectations could be adaptive rather than
forward-looking, or could be derived from a data-based (rather than a model-
based) predictor," "statistically acceptable cross-equation restrictions in expec-
tational models ... could be satisfied by a conditional-marginal representation as
well." With such statistical incantations, Ericsson and Irons reduce to two the
number of articles containing "fruitful" evidence about the Lucas critique. Ulti-
mately they reject these, too, on similar statistical grounds. Thus, exactly zero
papers find empirical support for the Lucas critique.
In contrast, the authors report that twenty-five papers find evidence for super
exogeneity and only five find evidence against it. According to Ericsson and Irons,
COMMENTARY ON CHAPTER 8 315
Ericsson and Irons embed their empirical tests in the now-standard rational expec-
tations framework for econometric policy evaluation, as articulated by Lucas (1976).
Private-sector behavior is summarized by
316 MACROECONOMETRICS
(1)
The function F and the parameter vector () are derived from dynamic optimizing
behavior on the part of private agents, so they represent decision rules. E, is a
vector of independent, identically distributed random shocks. Policy behavior is
characterized by
(2)
where G is a known function that, coupled with A, describes the decision rules of
the policy authorities, and 11, is a vector of independent, identically distributed
disturbances. Substituting equation (2) into (1) yields the model of the economy:
Y'+l = F(y" X" ()(A), E ,). (3)
The econometric problem is to estimate the function, ()(A), rather than a set of
fixed parameters, (). Policy evaluation, according to Lucas, consists of considering
the effects of various settings of 1.., not of forecasting y conditional on different
choices of time paths for x.
Ericsson and Irons, drawing on previous work by Hendry (1988) and others,
base their test of the empirical importance of the Lucas critique on two procedures:
For equation (1) the authors estimate a relationship whose dependent variable
is real money balances (m - p) and whose right variables include lagged real
balances, current and lagged inflation (ilp), the short-term interest rate (rs), the
long-term interest rate (rl), and a lagged error correction term (u) that depends on
real balances, real income (z), and the short rate:
2
il(m - p), = F(il(m - P),_I, il (m - P),-2, ilp" ilp'_l'
(4)
where u = m - p - z + a . RS + constant and D is a vector of dummy variables
for world wars and a regression constant. The authors report a final empirical spec-
ification for (4) that exhibits constant parameters. They interpret (4) as a money
demand function and the vector of estimated elasticities from (4) as () in (1).
The paper then shows that the stochastic processes for the three variables
that enter (4) contemporaneously (inflation, short and long interest rates) exhibit
COMMENTARYONC~R8 317
Ericsson and Irons have run up against the same logical problems that Sargent
(1984) discusses and the Cooley, leRoy, and Raymon (1984), and Sims (1986,
1987) emphasize. If it is reasonable to evaluate changes in policy that are repre-
sented as different settings of A, then it is unreasonable to model policy as if a
fixed A has been in place forever and to assume that private agents expect it to
remain fixed forevermore. S Sims has developed an alternative framework for policy
analysis that requires identifying policy behavior to avoid the logical problems
inherent in the rational expectations approach.
As Sims (1987) puts it, if A can be changed, then it is necessarily a time series,
and private agents fonn expectations over its future values. Subscripting A by time
and substituting (2) into (1) yields
y,+\ = F(y" G(y" A" 1],), 6(A,), E,) == F*(y" A" E~), (5)
where E~ == (E" 1],). F* has the same functional form as F with A, playing the role
of X,.6 Now the econometric problem is to estimate the unknown parameters in the
functional fonn of 6(A,). Sims's point is that there is no logical difference between
evaluating policy as changes in A or as alternative time paths for x. 7
This argument pertains to Ericsson and Irons's procedures, which rest er.tirely
on the identification that nonconstant marginal processes reflect changes in A, not
merely changes in the time paths of x. Different realizations of the policy dis-
turbance, 1]" and hence in x" do not generate changes in 6, according to the
rational expectations framework for policy evaluation. But if changes in A and
changes in x are not logically distinct, then according to the model economy in (3),
any sort of change in policy destabilizes private behavior, rendering all policy
evaluation meaningless. The alternative framework, however, explicitly models
the time-varying nature of policy and treats private agents as having probability
318 MACROECONOMET~CS
distributions over A. Under this framework, changes in Awithin some class do not
shift private decision rules so, according to (5), it is possible to conduct policy
analysis.
Ericsson and Irons pose policy changes as a grander thing than (5) envisions.
To execute the test that Ericsson and Irons want to perform, it is crucial to identify
the policy function G and periods when A changed in important ways. The test
consists of checking if changes in A are associated with changes in e. The only
way to perform this test is to identify the aspects of an economic model that
produce the F and G functions (and their associated parameter vectors).8
Solving the identification problem involves separating the private sector's re-
sponse to policy from policy's response to economic conditions. Otherwise one is
likely to make the mistake of interpreting endogenous policy reactions, which do
not engender changes in private decision rules, as shifts in policy that do.
Ericsson and Irons perform a battery of statistical tests from which they infer
that simultaneity is not a problem, so they do not need to model policy. This in-
ference requires one to believe that money demand disturbances have no effect on
short and long interest rates and prices within a year, which is the frequency of the
data Ericsson and Irons use. Policy can offset short-term interest rate movements
by fully accommodating demand shocks, but then the effects show up instead in
prices (possibly with a lag) and in long rates (likely immediately). Thus, Ericsson
and Irons's estimation of money demand embodies special and, I believe, implaus-
ible identifying assumptions about both policy behavior and the structure of the
rest of the economy.
Specifying and estimating the policy function, (2), also distinguishes policy
from nonpolicy disturbances. Ericsson and Irons interpret the instability of the
marginal processes for inflation and interest rates as stemming entirely from policy
behavior. This is a strong identifying assumption, as many factors unrelated to
policy can shift the stochastic structure of the economy.9 Although this point is not
important for evaluating the logic of the tests Ericsson and Irons perform, it is
crucial for interpreting their findings in the context of policy analysis. Whether a
change in the stochastic process for, say, technological innovation shifts the pro-
cesses for inflation and interest rates depends on how policy reacts to it. Essentially,
the function G in Ericsson and Irons's procedure is composed of a G I function
with parameters Al describing policy and a G2 function with parameters A2 de-
scribing all behavior other than policy and money demand. It is impossible to
discern from Ericsson and Irons's work whether changes in Al or A2 generated the
instability of the marginal processes.
COMMENTARY ON CHAPTER 8 319
Ericsson and Irons also maintain the assumption that any instability in the
marginal processes should generate instability in (4) if expectational behavior is
important. A more moderate view is that "big" changes in the marginal processes
should shift private behavior but that small ones will have more modest effects
over short horizons, even if individuals form expectations rationally. Because they
have confined themselves to the rational expectations framework Lucas laid
out, Ericsson and Irons's methodology treats all statistically significant changes
in the marginal processes as equally important. But whether a statistically signifi-
cant change is economically important depends on the economic model under
consideration.
The authors asset that "if the Lucas critique is refuted by tests 1 and 2, the
refutation is generic." But the authors do not discuss in detail, much less fully
specify, even a single economic model that implies the equations they estimate. So
exactly what are the authors refuting? This section explores some possible inter-
pretations of Ericsson and Irons's findings.
Without an economic model, the identification of (4) as a money demand
function is tenuous. Ericsson and Irons's final empirical specification raises sev-
eral questions that need to be addressed before many readers are willing to inter-
pret (4) as a behavioral relationship:
likely to be relatively stable over time. General equilibrium models that Lucas has
in mind imply demand functions of the fonn
function, like Ericsson and Irons's, passed the requisite battery of statistical tests,
including super exogeneity. Hess, Jones, and Porter added five years of quarterly
data to the original sample and found that the model's specification unambigu-
ously fails the authors' own statistical criteria. Hess, Jones, and Porter then present
evidence and a set of arguments that point toward the conclusion that the model's
failure is associated with overfitting.
Concluding Remarks
If the past twenty years of macroeconomic research have taught economists any-
thing, it has taught the dangers of fitting (and refitting) empirical specifications
that are not solidly grounded in economic theory. All the sophisticated econometric
techniques developed in recent years cannot substitute for sound economic reason-
ing and careful efforts to identify data correlations with economic behavior.
The main fault I find with Ericsson and Irons's work is that it frames policy
analysis in black and white terms when, in fact, policy analysis is largely gray.
Strategic interdependence of the sort that Lucas emphasizes certainly is present in
the actual economy. Its importance surely depends both on the private behavioral
relationship of interest and on the type of policy intervention being contemplated.
Even believers in the stability of money demand would be unlikely to argue that
the same functional form and demand elasticities would remain in effect if infla-
tion rose from 2 percent to 200 percent per year. Likewise, believers in the insta-
bility of the Phillips curve probably would not throw out a relationship estimated
using data that fluctuated mildly around 2 percent when inflation was contem-
plated to rise temporarily to 2112 percent.
Ericsson and Irons's methodology views policy analysis in its starkest form: a
relationship is stable or unstable; private behavior shifts with policy or it does not;
expectations affect behavior or they do not. Such starkness lends itself to statistical
testing but misses the subtlety inherent in actual policy analysis. By missing the
subtlety, the methodology loses its applicability for policy evaluation.
There is a continuum running from theoretically coherent models to ones that
fit data. This is an unfortunate reality. Theoretically coherent models can be un-
derstood well because they are simple; empirical models fit the data well because
they are not. One is always faced with this unpleasant tradeoff when trying to infer
economic behavior from data. Real business-cycle models following Kydland and
Prescott (1982), which are "calibrated" to match certain time series statistics, lie
on one end of the spectrum; LSE-style empirical work chooses to land at the other
end. Ericsson and Irons view the world through statistical glasses, while many
followers of Kydland and Prescott see things through the lens of economic theory.
Progress in empirical macroeconomics requires bifocals.
322 MACROECONOMETIUCS
Acknowledgments
Jon Faust, Kevin Hoover, and Will Roberds provided useful comments and
suggestions.
Notes
I. The alternative treatment among macroeconomists, of course, is to dismiss the critique as being
unimportant for their application.
2. Among the Cowles Commission writers on this topic were Marschak (1953), Koopmans and
Bausch (1959), and Hurwicz (1962).
3. They present no critical analysis of the eleven citations that do not support the Lucas critique.
4. Another blow against large-scale models was leveled by Sims (1980). Sims argued that the
identification of those models and, therefore, their economic interpretations are fragile. The reactions
to his criticisms have probably also been overdone.
5. Sargent admits the logical difficulties with the rational expectations approach to policy evalu-
ation. His solution is to treat policy as having been in a single regime historically and to assume that
policy behavior was not purposeful during the period. The assumptions about policy allow him to treat
changes in policy variables as identifying sources of variation off of which private parameters can be
estimated. Policy evaluation, according to Sargent, is a once-and-for-all exercise that consists of
solving for the parameters A that solve the policy authority's optimization problem.
6. In F* there are no time-invariant aspects of policy choice, but that can be modified easily.
7. Ericsson and Irons's remarks on Sims's views completely miss the point. Sims lays out a frame-
work within which conventional policy analysis, consisting of projecting the effects of alternative time
paths of policy variables, does not render unstable the decision rules of rational private agents. His
structure makes precise the class of policy interventions that will not destabilize the model.
8. Neft~i and Sargent (1978) and Hoover (1991) conduct tests along these lines.
9. Indeed, the real business-eycle literature launched by KydIand and Prescott (1982) questions
whether shifts in policy, especially monetary policy, have any important effects on the economy.
10. For example, Lucas's (1988) cash-in-advance model, McCallum's (1989) shopping-time model,
as well as a variety of money-in-the-utility function and transactions cost models. Richer models imply
demand specifications that depend on realized values of a vector of asset returns as well as some
measure of transactions or wealth, but not on their expected values. In contrast, the money demand
behavior emerging from the model in Gordon and Leeper (1995) depends explicitly on expected future
monetary and fiscal policy outcomes. Such a demand function shifts with the processes describing
policy behavior.
II. See Mil\er and Roberds (l99\), though, for an argument that Sims's econometric policy
evaluation procedures are not robust to big changes in policy.
12. A "structural" relationship need not connect the decision variable of a particular group of
economic decisionmakers to variables that they take as beyond their control. Conversely. a description
of the behavior of some group of decisionmakers need not be invariant or structural.
References
Baba, Yoshihisa, David F. Hendry, and Ross M. Starr. (1992). "The Demand for M1 in the
U.S.A., 1960-1988." Review of Economic Studies 59, 26-61.
COMMENTARY ON CHAPTER 8 323
Cooley, Thomas F., Stephen leRoy, and Neil Raymon. (1984). "Econometric Policy Evalu-
ation: Note." American Economic Review 74, 467-70.
Friedman, Milton. (1968). "The Role of Monetary Policy." American Economic Review 58.
1-17.
Goldfeld, Stephen. (1976). "The Case of the Missing Money." Brookings Papers on Eco-
nomic Activity 3. 683-730.
Gordon, David B., and Eric M. Leeper. (1995). "A Growth Model of Private Transactions
and Velocity." Manuscript, Federal Reserve Bank of Atlanta, February.
Hendry, David F. (1988). 'The Encompassing Implications of Feedback Versus FeedfolWard
Mechanisms in Econometrics." Oxford Economic Papers 40, 132-149.
Hess, Gregory D., Christopher S. Jones, and Richard D. Porter. (1994). 'The Predictive
Failure of the Baba, Hendry, and Starr Model of the Demand for MI in the United
States." Finance and Economics Discussion Series No. 94-34, Federal Reserve Board,
November.
Hoover, Kevin D. (1991). "The Causal Direction Between Money and Prices: An Alter-
native Approach." Journal of Monetary Economics 27. 381-324.
Hurwicz, Leonid. (1962). "On the Structural Form of Interdependent Systems." In Ernest
Nagel, Patrick Suppes, and Alfred Tarski (eds.), Logic and Methodology in the Social
Sciences (pp. 232-239). Stanford: Stanford University Press.
Koopmans, Tjalling C., and Augustus F. Bausch. (1959). "Selected Topics in Economics
Involving Mathematical Reasoning." SIAM Review I. 138-148.
Kydland, Finn, and Edward C. Prescott. (1982). "Time to Build and Aggregate fluctua-
tions." Econometrica 50, 1345-1370.
Lucas, Robert E., Jr. (1976). "Econometric Policy Evaluation: A Critique." In Karl Brunner
and Allan H. Meltzer (005.), The Phillips Curve and Labor Markets, Carnegie-Rochester
Conference Series on Public Policy. Vol. I, Journal of Monetary Economics, supple-
mentary issue, 19-46.
Lucas. Robert E.• Jr. (1988). "Money Demand in the United States: A Quantitative Re-
view." In Karl Brunner and Bennett T, McCallum (eds.), Money, Cycles, and Exchange
Rates: Essays in Honor ofAllan H. Meltzer, Carnegie-Rochester Conference Series on
Public Policy (Vol. 29) (pp, 137-168). Amsterdam: North-Holland.
Marschak, Jacob. (1953). "Econometric Measurements for Policy and Prediction." In William
C. Hood and Tjalling C. Koopmans (eds.), Studies in Econometric Methods (pp. 1-26).
New York: Wiley.
McCallum, Bennett T. (1989). Monetary Economics: Theory and Policy. New York:
Macmillan.
Miller, Preston 1., and William T. Roberds. (1991). "The Quantitative Significance of the
Lucas Critique." Journal of Business and Economic Statistics 9, 361-387.
Neft~i, Salih, and Thomas J. Sargent. (1978). "A Little Bit of Evidence on the Natural Rate
Hypothesis for the U.S." Journal of Monetary Economics 4,315-319.
Phelps, Edmund S. et al. (1970). Microeconomic Foundations ofEmployment and Inflation
Theory. New York: Norton.
Sargent, Thomas 1. (1984). "Vector Autoregressions, Expectations. and Advice." American
Economic Review 74, 408-415.
Sims, Christopher A. (1980). "Macroeconomics and Reality." Econometrica 48, 1-48.
324 MACROECONOMETIUCS
Sims, Christopher A. (1986). "Are Forecasting Models Usable for Policy Analysis?" Quar-
terly Review (Federal Reserve Bank) Winter, 2-16.
Sims, Christopher A. (1987). "A Rational Expectations Framework for Short-Run Policy
Analysis." In William A. Barnett and Kenneth J. Singleton (eds.), New Approaches to
Monetary Economics (pp. 293-308). Cambridge: Cambridge University Press.
9 RATIONAL EXPECTATIONS AND
THE ECONOMIC CONSEQUENCES
OF CHANGES IN REGIME
James D. Hamilton
Many economic variables undergo episodes in which the behavior of the series
changes quite dramatically. Sometimes this is caused by events such as wars,
financial panics, and economic recessions. Abrupt departures from the historical
pattern can also be the result of deliberate policy actions taken by the government.
For example, if inflation has become a very serious problem, the government may
adopt radical changes to try to bring inflation down quickly. Other sources of
abrupt change can be the introduction of new taxes and the elimination of previous
government regulations.
How economic variables behave in times of dramatic change is an important
area of study. This essay surveys recent research on this topic, with a particular
focus on how major changes in the economic environment or government policy
can affect the behavior of rational economic agents. For concreteness we begin
with a detailed discussion of government policies designed to eliminate high
inflation. We will subsequently comment on the ways in which similar considera-
tions apply in a variety of different settings.
325
326 MACROECONOMETRICS
=
m, natural log of the public's money holdings at date t,
P, = natural log of the aggregate price level at date t,
y, = natural log of the level of aggregate real income at date t,
i, = nominal interest rate at date t (measured as a fraction of one).
E,p'+1 - p,.
The ex ante real interest rate is defined as the nominal interest rate (i,) minus
expected inflation. If r, denotes the ex ante real interest rate, then
i r = r, + (ErPI+1 - P,).
To complete the model we need a theory of how the public's expectations of in-
flation are formed. One approach is based on the principle of rational expectations.
328 MACROECONOMETItlCS
This approach was originally proposed by Muth (1961) and developed more fully
by Lucas (1973) and Sargent and Wallace (1973). The approach requires the
model builder to specify the statistical properties of the exogenous variables. In
this case such a specification might take the form of an equation describing m, in
terms of its own past values and some random variable representing changes
in money that are impossible to forecast. The model builder then specifies the in-
formation on which people in the economy are presumed to base their forecast of
endogenous variables, and conjectures a form that this forecast would take. For the
model of hyperinflation we would guess a rule that people might use to form a
forecast of the future price level, E,p,+!. The final step is to verify that the model
is internally consistent-that is, to verify that the conjectured forecasts are indeed
rational ones to make given the dynamic properties of the variables that are im-
plied by the model.
For example, suppose that the money-supply process is described by
Here g is a parameter that measures the average growth rate of the money supply.
The variable e'+1 represents changes in the money supply that are impossible to
forecast. We could posit that e'+1 is drawn from a N(O, (1;) distribution and is inde-
pendent of previous values of any variable in the economy. This money supply
process implies that the government is just as likely to increase the money supply
by more than g percent as it is to increase the money supply by less than g percent.
We might conjecture that since the money supply grows by g percent on aver-
age, it would be rational for the public to expect g percent inflation:
E,P'+l - P, ::: g.
Substituting this into the money demand function (9.3) implies that
p, :::: m, - c + ago (9.4)
It is easy to verify that this model is internally consistent. If p, is described by the
above equation, them Pt+1 should be as well:
P'+I ::: m/+! - c + ago
The actual inflation rate would them be
P'+I - P, :::: (m'+1 - c + ag) - (m, - c + ag)
:::: m/+ I - m,
::: g + e/+ I •
Since e'+1 is impossible to forecast and is equal to zero on average, it would be
perfectly rational for people in such an economy to expect g percent inflation, as
we conjectured.
CHANGES IN REGIME 329
II -ag+
lJ _
1+
lag
___-:1 ,
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 1920
TIme
Figure 9.1. Behavior of the price level under changes in regime that are regarded
as permanent
Consider next the effects of a change in policy. Suppose that historically the
economy has had zero money growth and stable prices (g = 0), but beginning at
date T. the government allows the money supply to grow rapidly (g > 0). Suppose
that up until date Tl the relation between prices and money is given by the pricing
equation (9.4) with g = 0,
for t < T I ,
whereas after date T1 the appropriate equation is that for the positive inflation
environment,
p, = m, - c + ag
The time path followed by prices (under the simplifying assumption that er+1 =
o for all t) is plotted in Figure 9.1. Note that not only does the change in policy
cause the average inflation rate to move from zero to g percent per year, the
330 MACROECONOMETIUCS
change also induces a one-time surge in prices at date T( in the amount ago This
one-time surge in prices is the result of the flight from money described earlier.
These results point to a difficulty with basing policy predictions on historical
correlations. Suppose we had collected data from the zero inflation regime, and
used it to estimate the relation between prices and money that held for that histori-
cal episode:
P, = m, - C.
As noted by Lucas (1976), this historical relation will not accurately predict the
consequences of changing the money growth. The reason is that the historical
relation is influenced by the expectations about government policy that people
held during the sample. If the government changes its policy, the nature of peo-
ple's forecasts may change, and this implies a different relation between prices
and money-namely,
P, = m, - c + ago
The earlier historical relation between money and prices, P, = m , - c, will
underpredict the level of prices under the new regime.
If the government were to reform its policy at some later date Tz and return
money growth to zero, the above reasoning applies in reverse. The reform would
be accompanied by a one-time drop in the price level in the amount -ag (again
see Figure 9.1). Alternatively, if the government could credibly convince the
public that future money growth was going to be controlled, at the end of the
inflation (date Tz) the government could implement a one-time surge in the money
supply, increasing mTz by the amount ag, without any inflationary consequences.
Although this scenario might seem farfectched, it in fact describes exactly what
seems to have happened at the the end of many historical hyperinflations. If con-
vincing, dramatic monetary reforms are adopted, such as commitment to a fixed
exchange rate and a clear institutional separation of the central bank from the fiscal
authority, governments often have been able to increase the money supply dra-
matically at the end of big inflations. This increase does not drive up prices
because it simply is meeting the increased demand for money that comes at the
end of a hyperinflation. On a more modest scale, some economists have suggested
that the U.S. Federal Reserve's demonstrated willingness to tolerate very high
interest rates in the early 1980s convinced the public of its seriousness about
bringing down inflation. The result was that a large increase in the money supply
after 1982 produced little increase in inflation. Auernheimer (1974) argued that
analyses that leave out this potential surge in the money supply at the end of a
hyperinflation often underestimate the revenue benefits that a government can
enjoy by moving to a money supply path with a slower average growth rate.
CHANGES IN REGIME 331
While this framework offers some useful insights into hyperinflation, it is not
altogether satisfactory for analyzing the dynamic transition between low and high
inflation episodes. As emphasized by Cooley, LeRoy, and Raymon (1984), the
above analysis is not really true to the principle of rational expectations. Prior to
date T1, the people in the economy were assumed to attach zero probability to the
likelihood of an acceleration of money growth-that is, they regarded g 0 as a=
fixed constant. Yet as events turn out, the parameter g is evidently subject to
change.
A model closer in spirt to the ideas of Muth (1961) would carry the analysis
one step further and specify a probability distribution governing the value of the
average money growth parameter g. The simplest such model might postulate that
if the government is currently follOWing the zero inflation regime, there is a prob-
ability q that it will continue with the zero inflation regime the next period, and
a probability I - q that it will shift to the regime in which money grows at g per-
cent per year. That is, if the economy is in the zero inflation regime at date t, then
with probability q,
E,m'+1 = m, + (I - q)g.
Similarly, if the economy is in the high inflation regime at date t, there is a
probability w that it will continue in the high inflation regime next period, and a
probability (1 - w) that it will return to the zero inflation regime:
ag(l- q) vg
P, =-c+ +m, +---5" (9.5)
(1 - v) (1 - v)
where
a(-I+w+q)
v= .
(l + a)
The appendix verifies that equation (9.5) satisfies the internal consistency require-
ment of rational expectations.
If the economy is currently in the zero inflation regime (5, = 0), then as the
probability of remaining in that regime goes to one (q ~ 1), equation (9.5) ap-
proaches
P, =-c + m"
which is the earlier equation used for the pure zero inflation economy. Likewise,
if the economy is currently in the high inflation regime (5, = 1), then as the
probability of staying with high inflation goes to one (w ~ 1), the price process
can be shown to converge to
P, = -c + m, + ag,
which is the earlier equation used for the pure high inflation economy. Thus as
noted by Lucas (1976), the simple model employed in Figure 9.1 can be viewed
as a limiting case of a more general framework, the limiting case being one in
which people believe that each regime is virtually certain to persist forever. The
more general framework also applies to situations in which people's anticipations
of future changes in regime can materially affect the current equilibrium.
Suppose we have data that may have been characterized by changes in regime.
Even though the econometrician may not know when the changes occurred, the
observed data can be used to estimate the parameters describing each regime and
to form an inference about the most likely dates of changes in regime.
For example, consider a data set consisting only of money growth rates for a
series of different dates. Suppose that the term e 1+1 described above is drawn from
a N(O, <T.) distribution. If the economy is in the high inflation regime at date t, the
CHANGES IN REGIME 333
f(x)
probability that the economy will be in the high inflation regime at date t + I is
w. If the economy is in the zero inflation regime at date t, the probability that the
economy will be in the high inflation regime at date t + I is I - q. Consider the
simplest case in which the probability of being in the high inflation regime next
period is the same regardless of the current regime (w = I - q). Then a fraction
w of the observed money growth changes would likely be drawn from a normal
distribution with mean g, and the remaining fraction 1- w would be drawn from
a normal distribution with mean zero:
The density of (m1+ 1 - m,) is then a weighted average of the two normal densities:
!(m,+] - m t ; g, 0'., w) =
w
~ exp
{-(m m
t +) -
2
t - g)2 }
v2nO', 20',
see Everitt and Hand (1981, pp. 28-29) for other possible shapes this distribution
might assume.
Given data on money growth rates (mt+l - m,) for some sample of size T(t =
I, 2, ... , T), the values of the parameters g, 0'., and w can be estimated by the
principle of maximum likelihood. This principle suggests that we use as estimates
of g, 0'.. and w those values that would have been most likely to have generated
the observed data. Loosely speaking, the maximum likelihood estimate wwill
essentially be the fraction of observations that cluster in a bell-shaped curve
centered at some point above zero. The estimate g will be the sample mean of this
subset of the observations, while 0'; will be based on the average squared deviation
of each observation from its nearest mean.!
Once we have estimated the population parameters g, 0'., and w, these can be
used to form an inference about which regime was in effect for any given date in
the sample. The probability that 5'+1 = I-that is, the probability that money
growth between periods t and t + I is characterized by the high inflation regime-
is given by w. Moreover, if the economy is in regime I at date t + I, then (m'+1
- mt) is distributed N(g, 0';). Hence the joint probability that (1) 5'+1 = I and (2)
money growth would equal the observed value m t+1 - m t is given by
This quantity tells us how likely it is that the government was actually following
a high inflation policy between dates t and t + l.
The above description assumed that the current regime (5 t = 0 or I) has no effect
on the probability of next period's regime and that the term e'+1 is independent and
identically distributed across dates. However, the same basic idea can be applied
with more general dynamic behavior for both the regime and the error, and is
easily extended beyond two possible states (see Hamilton, 1989, 1990, 1994, for
details). The maximum likelihood estimate of w essentially turns out to be the
fraction of times that regime I appeared to be followed by another observation
from regime I, while the maximum likelihood estimate of q is roughly the fraction
of times that regime 0 was followed by another observation from regime O. Gen-
CHANGES IN REGIME 335
eral dynamic relations can be fit separately to the subset of observations that
describe each of the two regimes.
One might think that if a switch in regime only appears to have happened once
in a given sample, no reliable inference about the parameters of the process can
be drawn. However, this is not the case. The reliability of the inference about the
parameters that describe regime 1 depends on the total number of observations
that were generated from regime 1 and not on the number of switches between
regimes that occur. Every observation generated from regime 1 gives new infor-
mation about the parameter w; if regime 1 is almost always followed by another
observation from regime I, that is evidence that w is large. If the economy seems
to have remained in regime I for ten consecutive periods, for example, then we
can assert with some confidence that w must be larger than 0.7. 2
ag(l - q) vg
p, = -c + + m, + - - s , + u,
(I - v) (l - v)
m, ]
y, = [ p, - .
m, -m'-l
Then the model can be expressed as
y, - N«(.lo, I) when s, = 0,
y, - N«(.ll' I) when s, = I
where
I = [~ :;].
The public's perceptions of current and future changes in regime can have impor-
tant economic consequences. This possibility was illustrated above in terms of a
CHANGES IN REGIME 337
10
35
30
z 25
g
I- 20
<
...J
"- IS
i?;
10
0
83 81 85 86 87 88 89
1.00
0.75
111 0.50
ffi
0
t¥ 0.25
0-
0.00
83 81 85 86 87 88 89
Figure 9.3. Top panel: Monthly inflation rate in Brazil, 1983-1989; bottom panel:
Probability that Brazilian fiscal policy is in the high inflation regime
Source: Ruge-Murcia (1992).
One important example can occur if a government is trying to maintain a fixed rate
of exchange between its currency and some other country's currency. For exam-
ple, the Mexican government might try to peg its currency (the peso) to the U.S.
dollar. It does this by maintaining reserves of dollar-denominated assets. When the
public wants to sell pesos in exchange for dollars, the government provides a
ready demand for these pesos by buying these pesos with its dollar reserves. Such
purchases can keep the peso's value from falling.
338 MACROECONOMETIUCS
800
~v
....... /
V
780
760
710
/'
/ r-~
720
52 55 58 61 61 67 70 73 76 79 82
1.00 ,....
f1 ~
0.75 ~
l/) 050
U
1D
a
U
0<
c- 0.25
52 55 58 61 61 67 70 73 76 79 82
Figure 9.4. Top panel: 100 times the log of U.S. real GNP, 1951-1984; bottom
panel: Probability that the U.S. economy is in the recessionary regime
Source: Hamilton (1989).
The top panel of Figure 9.4 graphs the log of quarterly real GNP for the United
States over 1951-1984. The general upward trend is occasionally interrupted by
short episodes in which GNP actually falls, known as economic recessions. The
dates at which the National Bureau of Economic Research believes that recessions
began and ended are indicated with vertical lines on the top panel. These dates are
based on ex post judgmental examination of a number of economic indicators.
When GNP growth rates are viewed as having come from two different dy-
namic regimes, the maximum likelihood estimates suggest that one of the regimes
is associated with rising GNP and the other with falling GNP. The bottom panel
of Figure 9.4 (taken from Hamilton, 1989) graphs the probability that the economy
was in the faIling GNP regime at any given historical date. This designation is
remarkably similar to that arrived at by the National Bureau of Economic Re-
search based on qualitative judgment.
340 MACROECONOMETRICS
Summary
The dynamic behavior of many economic time series can sometimes change dra-
matically from the previous historical pattern. The public's expectations about
future changes have important implications for how the economy behaves today.
Using a statistical model to describe these changes in regime offers a useful
approach for analyzing their effects on rational economic agents.
CHANGES IN REGIME 341
£ - --g-£(s )-~s
IPI+I PI - (I _ v) I 1+1 (1- v) I'
Recalling that
it follows that
E _ _ g(l-q) +g(-I+w+q)s -~s
IPt+l PI - (I - v) (I - v) I (I - v) I
= g(1 - q)
(I - v)
g
+ --(-I + w + q -
(I - v)
V)SI' (9A.l)
But since
a(-I + w + q)
-I + w + q - v =(-I + w + q) - --'------''-
(I + a)
= {_(l_+_a_) __ a_
(I + a) (I + a)
}(-I + w + q)
(-I+w+q)
=
(l + a)
expression (9A.1) becomes
E _ = 8(1 - q) + 8(-1 + w + q) S
,P'+l P, (I - v) (I - v)(1 + a) I'
342 MACROECONOMETIUCS
p, = m, - c + a(E,p,+\ - p,)
ag(1 - q) ag(-l + w + q)
=m, - c++
(I - v) (1 - v)(1 + a)
s,
ag(l- q) vg
= m, - c + + - - s ,
(I - v) (1 - v)
as claimed in (9.5).
In the special case when s, = 1 this implies
g[a(1 - q) + v]
p, = m, - c+
(1- v)
. (9A.2)
_ a{l _ (I + a) +
- (1 + a) q
_q_}
(I + a)
= a{I-~}
(1 + a)
= a(l - v),
ag(1- v)
p, = m, - c+ -=:,,_-,-
(I - v)
= m, - c + ag,
reproducing the earlier result (9.4) for the pure high inflation regime as a special
case.
Acknowledgments
Notes
I. See Everitt and Hand (1981, pp. 36-37) for a more formal characterization of the maximum
likelihood estimates.
2. This statement is based on the observation that 0.7 10 = 0.028, which means the probability of
observing such a run if the true value of w were 0.7 or smaller would be less than 3 percent.
3. This argument is developed in Hamilton (l988b).
References
Auemheimer, Leonardo. (1974). ''The Honest Government's Guide to the Revenue from
the Creation of Money." Journal of Political Economy 82, 598-606.
Bemanke, Ben, and Harold James. (1991). "The Gold Standard, Deflation, and Financial
Crisis in the Great Depression: An International Comparison." In R. Glenn Hubbard
(ed.), Financial Markets and Financial Crises (pp. 33-68). Chicago: University of
Chicago Press.
Cagan, Phillip. (1956). "The Monetary Dynamics of Hyperinflation." In Milton Friedman
(ed.), Studies in the Quantity Theory of Money (pp. 25-117). Chicago: University of
Chicago Press.
Ceccheni, Stephen G., Pok-sang Lam, and Nelson C. Mark. (1990a). "Evaluating Empirical
Tests of Asset Pricing Models: Alternative Interpretations." American Economic Re-
view 80, 48-51.
Cecchetti, Stephen G., Pok-sang Lam, and Nelson C. Mark. (1990b). "Mean Reversion in
Equilibrium Asset Prices." American Economic Review 80,398-418.
Cooley, Thomas E, Stephen F. leRoy, and Neil Raymon. (1984). "Econometric Policy
Evaluation: Note." American Econamic Review 74, 467-470.
Eichengreen, Barry, and Jeffrey Sachs. (1985). "Exchange Rates and Economic Recovery
in the 1930s." Journal of Econamic History 45, 925-946.
Engel, Charles M., and James D. Hamilton. (1990). "Long Swings in the Dollar: Are They
in the Data and Do Markets Know It?" American Economic Review 80, 689-713.
Everitt, B.S., and D.J. Hand. (1981). Finite Mixture Distributions. New York: Chapman
and Hall.
Filardo, Andrew. (1994). "Business Cycle Phases and Their Transitional Dynamics." Jour-
nal of Business and Economic Statistics 12, 299-308.
Flood, Robert P., and Peter M. Garber. (1984). "Gold Monetization and Gold Discipline."
Journal of Political Economy 92, 90-107.
Friedman, Milton, and Anna J. Schwartz. (1963). A Monetary History ofthe United States.
1867-1960. Princeton, NJ: Princeton University Press.
Garcia, Rene, and Pierre Perron. (1989). "An Analysis of the Real Interest Rate Under
Regime Shifts." Mimeo, University of Montreal.
Hamilton, James D. (1988a). "Rational-Expectations Econometric Analysis of Changes in
Regime: An Investigation of the Term Structure oflnterest Rates." Journal ofEconomic
Dynamics and Control 12, 385-423.
344 MACROECONOMETRICS
Hamilton, James D. (l988b). "The Role of the International Gold Standard in Propagating
the Great Depression." Contemporary Policy Issues 6, 67-89.
Hamilton, James D. (1989). "A New Approach to the Economic Analysis of Nonstationary
Time Series and the Business Cycle." Econometrica 57, 357-384.
Hamilton, James D. (1990). "Analysis of Time Series Subject to Changes in Regime."
Journal of Econometrics 45,39-70.
Hamilton, James D. (1994). Time Series Analysis. Princeton NJ: Princeton University
Press.
Kindleberger, Charles P. (1989). Manias, Panics, and Crashes: A History of Financial
Crises (2nd ed.). New York: Basic Books.
Krugman, Paul. (1979). "A Model of Balance-of-Payments Crises." Journal of Money,
Credit, and Banking 1,311-325.
Lam, Pok-sang. (1990). "The Hamilton Model with a General Autoregressive Component:
Estimation and Comparison with Other Models of Economic Time Series." Journal of
Monetary Economics 26, 409-432.
Lucas, Robert E., Jr. (1973). "Some International Evidence on Output-Inflation Trade-
Offs." American Economic Review 63,326-334.
Lucas, Robert E., Jr. (1976). "Econometric Policy Evaluation: A Critique." In Karl Brunner
and Allen H. Meltzer (eds.), The Phillips Curve and Labor Markets. Vol. I, Carnegie-
Rochester Conference Series on Public Policy, 19-46.
Mishkin, Frederic S. (1991). "Asymmetric Information and Financial Crises: A Historical
Perspective." In R. Glenn Hubbard (ed.), Financial Markets and Financial Crises (pp.
69-108). Chicago: University of Chicago Press.
Muth, John F. (1961). "Rational Expectations and the Theory of Price Movements."
Econometrica 29, 315-335.
Phillips, Kerk L. (1991). "A Two-Country Model of Stochastic Output with Changes in
Regime," Journal of International Economics 31, 121-142.
Rodrik, Dani. (1992). "The Limits of Trade Policy Reform in Developing Countries."
Journal of Economic Perspectives 6,87-105.
Ruge-Murcia, Francisco. (1992). "Government Credibility in Heterodox Stabilization Pro-
grams." Ph.D. dissertation, University of Virginia.
Sargent, Thomas 1., and Neil Wallace. (1973). "Rational Expectations and the Dynamics
of Hyperinflation." International Economic Review 14, 328-350.
Schwert, G. William. (1989). "Business Cycles, Financial Crises, and Stock Volatility." In
Karl Brunner and Allen H. Meltzer (eds.), IMF Policy Advice, Market Volatility, Com-
modity Price Rules, and Other Essays. Vol. 31, Carnegie-Rochester Conference Series
on Public Policy, Autumn 83-125.
Schwert, G. William. (1990). "Stock Volatility and the Crash of '87." Review of Financial
Studies 3, 77-102.
Turner, Christopher M., Richard Startz, and Charles R. Nelson. (1989). "A Markov Model
of Heteroskedasticity, Risk, and Leaming in the Stock Market." Journal of Financial
Economics 25, 3-22.
Commentary on Chapter 9
Glenn D. Rudebusch
• How do rational economic agents identify and react to major changes in their
economic environment?
• How do econometricians identify and conduct inference in the presence of
such structural changes?
• What are the causes and mechanisms of such structural changes?
345
346 MACROECONOMETRlCS
Regime-Switching Models
In Hamilton's example, the application of the Lucas critique implies that the
equation that appears to adequately describe rational agents' money demand dur-
ing a zero-inflation regime breaks down with a change in the policy regime. The
reduced form relationship between prices and money changes from one regime to
the other. What is unsatisfactory about this analysis, as pointed out by Hamilton
as well as Sims (1982) and fully developed by Cooley, LeRoy, and Raymon
(1984) and by LeRoy's work in this volume, is that it has the rational agents treat
the structural change as a change in a parameter. That is, the change is treated as
unforeseen and permanent. More plausibly in most cases, the object that changes
should be treated as a variable and incorporated in the model. Hamilton provides a
useful example of how a change in policy can be incorporated as a probabilistic ele-
ment in a larger model. In his model, the agent with rational, or model-consistent,
expectations correctly recognizes and takes into account the probability that the
policy regime may change.
Of course, not all potential structural changes can be fruitfully incorporated
into a complete, all-encompassing model that is understood by rational agents.
Indeed, an interesting literature has developed that continues to treat the structural
change as an unforeseen, permanent parameter shift. However, this literature is
sensible because it relaxes the assumption of the rationality of the economic agents
and introduces explicitly the notion that some learning about the new structure
must take place. For example, Sargent (1993) provides a useful introduction to the
literature and centers his analysis on the "bounded rationality" of agents with
respect to structural change. Thus, rather than treating a change in a parameter as
something rational agents searnIessly adapt to (as did Lucas), the literature with
learning permits agents to slowly adjust their behavior to new situations in which
their previous experience is not entirely informative. For example, understanding
the economic actions of those in the newly emerging market economies of Eastern
Europe, who face unprecedented changes, realistically requires an understanding
of transitional learning dynamics.
In any case, to return to Harnilton's analysis, where there is an explicit process
generating the structural change, the next question he addresses is how to conduct
an econometric analysis of the changes in regime. Hamilton's model in this paper
is based on his pioneering work examining series whose time-series dynamics
depend on an unobservable state that is governed by a first-order Markov process.
There have been numerous applications of this specific regime-switching model-
perhaps the best known is the use in a business-cycle context for dating turning
points between recessions and expansions. I
One shortcoming of almost all of these applications is that they make no at-
tempt to statistically test the hypothesis of regime switching against the hypothesis
348 MACROECONOMETRICS
of a constant structure. The infrequency of testing for regime switching reflects the
difficult, nonstandard econometrics involved. The difficulty lies in the Markov
probabilities governing the transition, which are not identified under the hypo-
thesis that there is no regime switching. Hansen (1992) proposes a valid but com-
putational burdensome test. Diebold and Rudebusch (1994) describe and implement
a closely related but tractable test, and they find striking support for a two-state
regime switching process governing business cycles.
Finally, with respect to "explaining" regime switching, Diebold and Rudebusch
(1994) survey some interesting models that can be construed as supporting regime
switching in a business-cycle context. These are models in which there is a
"strategic" element to an agent's economic actions. These strategic elements or
externalities can arise, for example, in a model of search; in essence, search is
more desirable when other agents are also searching because it is likely to be
more productive. These externalities can produce multiple equilibria, and the
dynamic transition between these equilibria may be fruitfully modeled by a
regime-switching model.
Acknowledgments
The views expressed are those of the author and are not necessarily shared by the
Federal reserve Bank of San Francisco or the Federal Reserve System.
Note
1. Note that Hamilton's process can be used in real time by an agent to discern turning points. A
different real-time procedure to uncover turning points that also treats expansions and contraction as
different probabilistic objects is analyzed in Diebold and Rudebusch (1989, 1991).
References
Cooley, T., S. leRoy, and N. Raymon. (1984). "Econometric Policy Evaluation: Note."
American Economic Review 74, 467-470.
Diebold, F., and G. Rudebusch. (1989). "Scoring the Leading Indicators." Journal ofBusi-
ness 62, July, 369-391.
Diebold, F., and G. Rudebusch. (1991). "Turning Point Prediction With the Composite
Leading Index: An Ex Ante Analysis." In K. Lahiri and G.H. Moore (eds.), Leading
Economic Indicators: New Approaches and Forecasting Records (pp. 231-256). Cam-
bridge: Cambridge University Press.
Diebold, F., and G. Rudebusch. (1994). "Business Cycles: A Modem Perspective." Review
of Economics and Statistics, forthcoming.
Dufour, J. (1982). "Recursive Stability Analysis of Linear Regression Relationships." Journal
of Econometrics 19,31-76.
Hall, R. (1978). "Stochastic Implication of the Life Cycle-Pennanent Income Hypothesis:
Theory and Evidence." Journal of Political Economy 86, 971-987.
Hansen, B. (1992). "The Likelihood Ratio Test Under Non-Standard Conditions: Testing
the Markov Trend Model of GNP." Journal of Applied Econometrics 7, S61-S82.
Lucas, Robert E. (1976). "Econometric Policy Evaluation: A Critique." In K. Brunner and
A.H. Meltzer (eds.), The Phillips Curve and Labor Markets. Carnegie-Rochester Con-
ference Series on Public Policy (Vol. 1) (pp. 19-46). Amsterdam: North-Holland.
Marschak,1. (1953). "Econometric Measurements for Policy and Prediction." In William
C. Hook and TjaIling G. Koopmans (eds.), Studies in Econometric Method (pp. 1-26).
Cowles Commission Monograph 14. New York: Wiley.
Oliner, S., G. Rudebusch, and D. Sichel. (1995). "New and Old Models of Business invest-
ment: A Comparison of Forecasting Perfonnance." Journal of Money, Credit, and
Banking, forthcoming.
Oliner, S., G. Rudebusch, and D. Sichel. (1996). "The Lucas Critique Revisited: Assessing
the Stability of Empirical Euler Equations for Investment." Journal of Econometrics,
forthcoming.
Sargent, T. (1993). Bounded Rationality in Macroeconomics. Oxford: Clarendon Press.
Sims, C. (1982). "Policy Analysis With Econometric Models." Brookings Papers on Eco-
nomic Activity I, 107-152.
Tinbergen, J. (1956). Economic Policy: Principles and Design. Amsterdam: North-
Holland.
HISTORICAL 10
MACROECONOMICS AND AMERICAN
MACROECONOMIC HISTORY
Charles W. Calomiris
Christopher Hanes
Introduction
351
352 MACROECONOMETRlCS
Our focus in this essay is to highlight the factual, theoretical, and econometric
implications of historical research on American business cycles.
Sometimes the past has been used simply to extend time-series data to permit
tests of questions of current interest. The data and methods of such studies will
change over time as new paradigms are introduced by macroeconomists. Perhaps
the most famous and influential example of this approach is Friedman and
Schwartz's (1963) Monetary History o/the United States. This was arguably the
first and most influential test of the monetarist proposition that money and nominal
income are closely related. The authors' argument was based largely on the
robustness of the association over a long span of time.! More recently, macro-
economists in search of longer time series have turned to historical data to resolve
debates between the "new Keynesian" and "new classical" schools, which often
translate into debates about the cyclical properties of wages, prices, and tech-
nological change.
In some cases, macroeconomists have been drawn to particular data from his-
torical periods because those data are of better quality than any postwar data on
the same subject-for example, surveys of wage rate changes from the 1890s and
1930s (as we discuss below). More important, some interwar and prewar events
may constitute shocks to the economy that are more dramatic, and possibly more
informative, than any shocks that occurred in the postwar period. The Great
Depression is a case in point. As Milton Friedman noted long ago, "The major
defect of the data on which economists must rely~ata generated by experience
rather than deliberately contrived experiment-is the small range of variation they
encompass" (1952, p. 612).
The past can also be useful as a source of unique "historical experiments" that
can shed light on questions of current interest to macrotheorists or policy makers.
For example, Calomiris and Hubbard (1994a) use observed marginal tax rates paid
by firms subject to the short-lived undistributed profits tax of 1936-1937 (which
taxed retained earnings with a progressive marginal tax schedule) to measure
the shadow cost of external finance to firms and to show how differences in
the shadow cost of finance affect investment behavior. Some of history's natural
experiments take on new importance in light of later experience and policy issues.
The high-inflation episodes of the 1970s renewed interest in past experiences with
hyperinflation, and in the question of whether policy regime changes in the past
that ended high inflation were associated with major contractions in real activity
(Sargent, 1981a, 1981b). During the 1980s, controversies over rising federal defi-
cits prompted new research on extreme historical examples of high deficits in the
past and their links to inflation and interest rates (Smith, 1985a, 1985b; Calomiris,
1988a, 1988b; Evans, 1985; Barro, 1987).
But we will argue that macroeconomic history offers more than long strings of
data and special examples. It suggests an historical definition of the macroeconomy,
HISTORICAL MACROECONOMETRICS 353
variety of purposes-to increase sample size for testing models, to address ques-
tions of current interest with unique experiments from the past, and to demonstrate
the uses of the historical view of the economy (in particular, the importance of
institutional change). Second, we emphasize and draw out methodological issues
that arise in these examples and their implications for econometric modeling.
Our review of American macroeconomic history begins with an overview of
the salient facts about the U.S. macroeconomy during the nineteenth and twentieth
centuries, which focusses on continuity and change in patterns of covariation
among macroeconomic time series. The next section discusses the uses of these
and other data for discriminating among models of high-frequency macroeconomic
fluctuations. We focus on uses of historical data to distinguish between "new Key-
nesian" and "new classical" paradigms and to distinguish among new Keynesian
explanations for wage and price rigidities. We argue that Keynesian models can
account for the most obvious cyclical patterns in macroeconomic variables in all
historical periods, while new classical models cannot. New classical models are
also inconsistent with historical evidence on the nature of technological change.
It seems unlikely, therefore, that historical data will prove to offer much support
for real business-cycle theory, at least in its current form. We also argue that
nominal wage rigidity (which has been deemphasized by macroeconomists in
response to the "Barro critique") was important historically in generating layoffs
and that some models of nominal wage rigidity and its allocative costs receive
more support from history than others.
With respect to low-frequency fluctuations, neither the new-classical emphasis
on exogenous technological change nor Keynesian nominal rigidities offers a
complete account. Here historical evidence can be particularly useful for defining
the interesting unit of observation (the Kuznets cycle) and pointing to plausible
explanations for "long swings" by focusing on common historical patterns (city
building, frontier settlement, waves of immigration, and endogenous technologi-
cal change).
The final two sections discuss evidence of the importance of institutions, regu-
lations, and beliefs for macroeconomic outcomes, consider how important changes
in these variables have occurred (using the Great Depression as an example), and
offer conclusions.
From the standpoint of the methodological lessons we seek to draw out of our
review, each of these sections makes distinct but related points. The following
section focuses on problems of constructing consistent time series over long
periods of time, and analyzing those data in the presence of shifting historical
regimes (particularly, shifts in the time series process of prices). The next two
sections criticize both prominent macroeconometric frameworks for ignoring sets
of facts inconsistent with their models. In particular, we emphasize the failure
HISTORICAL MACROECONOMETRICS 355
As of ten years ago, most research had indicated two big historical changes in
cyclical patterns. First, output and employment were much less volatile after
Wodd War II than before Wodd War I or in the interwar period. Second, nominal
wages and prices were less "flexible" in the postwar period than in earlier periods,
in the sense that wage and price inflation had become less sensitive to deviations
from trend in output and employment levels. This evidence was cited to support
theoretical conclusions and even recommendations for policy (for an example see
DeLong and Summers, 1986). Since then careful attention to the nature of histori-
cal time series has somewhat modified these stylized facts.
356 MACROECONOMETRICS
Comparisons between historical periods require series that are consistent across
the periods: constructed in the same way from the same kinds of data. Otherwise
one may mistake differences in the way the series were constructed for changes
over time in economic behavior. If consistent data are not available, one must take
account of the inconsistencies and adjust one's conclusions accordingly. Econo-
mists largely ignored this point until Christina Romer (1986a, 1986b) asserted that
standard output and employment series for the late nineteenth and early twentieth
centuries had been constructed in ways that "exaggerated" cyclical movements.
Romer accepted the conventional wisdom that the Great Depression had been
much more severe than any postwar cycle, but she argued that otherwise "the
relative stabilization of the postwar economy is a figment of the data.... the
severity of economic fluctuations on both sides of the Great Depression are roughly
equal" (1986b, p. 333).
Others have challenged Romer's results (for a review, see Zarnowitz, 1992, ch.
3). David Weir (1986, 1992) concluded: "To the simple question of whether
cyclical fluctuations around trend in GNP and unemployment have become smaller
since World War II the data are more than adequate to deliver a definitive answer:
yes" (1986, p. 365). Nathan Balke and Robert Gordon (1989) constructed historical
series for real GNP that indicate "substantially greater volatility before 1929 than
after 1946" (p. 81). The standard deviation of deviation from trend in the log of
the series is about 70 percent greater over 1869-1908 or 1869-1928 than in post-
war periods. Indeed, Romer's own prewar real GNP series (1989) is more volatile
than the postwar series, with a standard deviation of deviation from trend that is
30 percent greater over 1869-1928 than 1947-1985. In Romer's view this means
the prewar series is "only slightly more volatile" than the postwar data (p. 35).
Others might judge the difference to indicate an important stabilization of output.
Unfortunately, no employment or GNP series for prewar years can be truly
comparable to postwar series. Annual data on many sectors of the economy were
simply not collected before the 1920s. Historical estimates must be built on debat-
able assumptions. In the words of Susan Carter and Richard Sutch (1990), the pro-
cess is like "inferring the shape of some long-extinct animal from bones collected
in an ancient tar pit" (p. 294). It appears unlikely that we will reach a consensus
any time soon. One might put more faith in results using data that are more con-
sistent and directly comparable, if less comprehensive. Romer (1986b) compares
the behavior of the best prewar series on industrial production, Edwin Frickey's
index of manufacturing output (1947), to a couple of very similar postwar series
on industrial production. She finds the standard deviation of deviation from trend
to be 10 to 14 percent greater over 1866-1914 than 1947-1982-a clear decrease
in volatility but smaller than that indicated by anyone's real GNP series.
HISTORICAL MACROECONOMETRICS 357
The difference between results from series on real GNP and industrial pro-
duction raises an issue that has been more or less ignored in the literature. Simon
Kuznets (1951) observed that sectors vary in their sensitivity to business cycles.
The cyclical behavior of aggregate output and employment might change over
time as a result of shifts in the relative importance of different sectors, even if
"there are no marked secular shifts within each sector in responsiveness to busi-
ness cycles.... For example, a decline in the weight of agriculture, combined
with a lack of responsiveness of agricultural output to business cycles, would
mean, other conditions being equal, a widening of business cycle amplitudes"
(p. 159). It made sense to focus on agriculture, for National Bureau researchers
like Wesley Mitchell and Arthur Bums had found farming to be a uniquely acyclical
sector: "the basic industry of growing crops does not expand and contract in
unison with mining, manufacturing, trading, transportation, and finance" (Mitchell,
1951, p. 56). Farm output and employment "undergo cyclical movements, but they
have little or no relation to business cycles" (Bums, 1951, pp. 7, 8). This was
especially obvious in the interwar period. From 1930 to 1932 employment fell in
every nonagricultural sector, including trade and services. Aggregate employment
fell by 14 percent. Meanwhile farm employment increased by 3 percent. 4 But it
was also true before World War I. Edwin Frickey (1942, p. 229) found that from
the 1870s through 1914 "agricultural production patterns traced out short-term
fluctuations bearing little resemblance to those for other major production groups.
The causal relationships between the agricultural and non-agricultural groups
certainly did not express themselves in the form of any simple correlation." This
is not to say that agricultural incomes are acyclical; that depends on the cyclical
pattern in the relative price of farm output.
As Table 10.1 shows, the twentieth century saw an enormous shift of resources
out of agriculture, while the share of manufacturing in employment or output did
not fall until the 1970s. Most of the balance went to services. Services output and
employment are less cyclical than manufacturing but more cyclical than agricul-
ture. By itself, the sectoral shift should have tended to decrease the cyclicality of
nonagricultural GNP but increase the cyclicality of real GNP including agricul-
ture. Both the Romer and Balke and Gordon series for real GNP purport to meas-
ure the latter. It seems odd that both indicate a larger decline in volatility than the
consistent manufacturing production series examined by Romer.
Suppose, however, that one had reliable historical series for both real GNP and
manufacturing output. Which would be appropriate to indicate changes in cyclical
behavior? In comparing cyclical "volatility" over long periods, should one hold
sectoral weights fixed or, equivalently, look at individual sectors? This issue is
especially important for comparisons of cyclical volatility across different periods
within the nineteenth century, before and after the War Between the States, for
example. John James (1993) found that Robert Gallman's unpublished real GNP
358 MACROECONOMETRlCS
Output Employment
% of GNP % of total labor force
Farm Manufacturing
Census year gross product value added Agriculture Manufacturing
1840 47 16 63 9
1850 37 19 55 15
1860 36 20 53 14
1870 31 20 53 19
1880 32 21 51 19
1890 22 27 43 19
1900 21 28 40 20
1910 31 22
1920 26 27
1930 9 30 22 20
1940 7 27 17 20
1950 7 29 12 24
1960 4 33 8 23
1970 3 33 4 23
1980 2 27 3 19
1990 2 25 3 15
Sources: Employment and labor force 1840-1960 from Lebergott (1964. table I). Employment and
labor force 1970-1990 from U.S. Council of Economic Advisers (1993). Farm gross product 1840-
1900 from Towne and Rasmussen (1960). Manufacturing value-added 1839-1859 from Gallman
(1960. p. 43). GNP 1839-1859 from Gallman (1966. p. 26); 1869-1899 from Balke and Gordon (1989,
table 10). All other series from U.S. Bureau of the Census (1975): GNP 1929-1969 series F1; farm
gross product 1929-1969 series K220-239; manufacturing value-added 1929-1969 series PI-12.
Remaining years from issues of Statistical Abstract of the United States.
series (described in Gallman, 1960) indicates "a substantial increase in the degree
of business-cycle severity or economic fluctuations" (p. 710) from 1834-1859 to
1869-1909. Calomiris and Hanes (1994) show that consistent series on industrial
production are if anything less volatile in the postbellum period. These two results
are not necessarily inconsistent with each other: given the rapid growth of manu-
facturing relative to agriculture over the nineteenth century, it is quite possible that
real GNP became more volatile even as manufacturing output became less volatile.
Historical changes in the timing of business cycles-duration from peak to
trough, and trough to peak-have also been the subject of recent studies. Francis
Diebold and Glenn Rudebusch (1992, p. 1003) show that the standard NBER
chronology suggests "a postwar shift toward longer expansions and shorter con-
tractions" with "no evidence for a postwar shift in the distribution of whole-cycle
HISTORICAL MACROECONOMETRlCS 359
durations." Geoffrey Moore and Victor Zarnowitz (1986) made the same obser-
vation but noted that the comparison was tricky because the NBER peak-trough
dates for the prewar period were not chosen in the same way as interwar and
postwar dates. Mark Watson (1994) and Christina Romer (1994) present evidence
that the change in timing of NBER cycles indeed reflects changing definitions
rather than changing economic behavior. They argue that contractions and expan-
sions were of about the same length in both the prewar and postwar periods.
Overall, the conclusion to draw from these various studies seems to be that
business-cycle severity has fallen over time while the duration and frequency of
cycles has not significantly changed across historical eras, excluding the interwar
period. The extent of the decline in cycle variance depends on assumptions both
about how to measure output comparably over time and whether to focus on
manufacturing or the whole economy.
(4) Ratio, inflation change over change in output deviation, NBER downturns
the change in the rate of inflation and the output gap, and hence they argue their
results are inconsistent with the notion of a Phillips curve. In fact, that is not the
case. Specification (5) shows that the correlation between the change in inflation
and output deviation is in fact positive in both periods. The correlation is larger
in the later period-from a Phillips curve point of view, an artifact of constraining
the coefficient on lagged inflation to be one.
362 MACROECONOMETIUCS
A few studies have used changes in the Phillips curve to measure changes in
the covariation of output with wage and price changes within the prewar period,
rather than between the prewar and postwar periods. Hanes (1993) presents evid-
ence of a change around 1890: the Phillips curve coefficient on output is larger and
the coefficient on lagged inflation is smaller in the 1870s and I880s than the 1890s
and 1900s. Hanes's result is consistent with James's (1989) study of price behavior
over 1837 to 1909: "the American economy earlier characterized by rapid (or in-
stantaneous) price adjustment to clear markets evolved into one marked by gradual
price adjustment (and one in which markets do not clear instantaneously) in the
last part of the 19th century.... the movement toward increasingly sluggish price
adjustment was most pronounced between 1880 and 1890" (p. 126).
But historical patterns of output-price correlations over time are clearer to
document that to interpret. Despite the tendency in many studies to use Phillips
curve coefficients to measure wage and price rigidity, we argue below that this is
not a straightforward exercise. In particular, there are many pitfalls in inferring
wage- and price-setting behavior from Phillips curve coefficients and in using
Phillips curves to measure the unemployment consequences of wage and price
rigidity.
Backus and Kehoe (1992) measure historical patterns in a number of other variables
for the United States and some other countries, comparing prewar, interwar, and
postwar periods. For all countries, in all periods, consumption appears "uniformly
procyclical" and investment is "strongly procyclical." Net exports have "generally
been countercyclical"; for the United States net exports appear countercyclical in
all periods.
Bernanke and Powell (1986) compare the cyclical behavior of employment,
weekly hours, average hourly earnings, and physical labor productivity in a fixed
set of industries across the interwar and postwar periods. They find that "the inter-
relation of productivity, hours, output, and employment is essentially stable between
the prewar and postwar periods" (p. 597), though they do observe "an increased
reliance in the postwar period on layoffs, rather than short workweeks, as a means
of reducing labor input." The real consumption value of hourly earnings-an
imperfect measure of "real wages"-was "procyclical (essentially coincident)
in the postwar period but 'half out of phase' (usually lagging)" in the interwar
(p. 617). They find no change in the cyclical behavior of labor productivity. "Pro-
cyclical labor productivity ... appears to be present in every industry, in both ...
periods" (p. 617). The lack of comparable employment data makes it hard to say
whether labor productivity was also procyclical before 1919. Some evidence sug-
HISTORICAL MACROECONOMETRICS 363
gests that labor productivity was not procyclical over the depression of the 1890s
(Weir, 1986; Carter and Sutch, 1990).
Cyclical patterns in a few other variables are of particular interest for our dis-
cussion below. One is the rate at which a domestic producer could exchange
output for foreign goods-that is, the terms of international trade net of domestic
tariffs. Robert Lipsey (1963) presents an annual series of terms of trade for the
United States beginning in 1879. Unfortunately the series is not a true index; it is
based on values per unit in fairly broad categories of goods and thus affected
by variations in the composition of goods as well as prices (p. 93). In the postwar
period terms of trade excluding tariffs can be roughly measured by the GNP de-
flator for exports divided by that for imports; these series begin with 1959. Tariff
rates are even more of a problem, since the usual measure, tariff revenue divided
by the value of imports, varies with shifts among commodities as well as changes
in rates. Indices of tariff rates for interwar years have been constructed by E.
Lerdau (1957) and Mario Crucini (1994). Crucini presents indices with both fixed
and changing commodity weights. There are no tariff rate series for earlier years,
but we may at least note that Frank Taussig (1931) describes no general rate
changes preceding prewar downturns.
We regress the terms of trade gross or net of tariffs on industrial production,
with both expressed as deviations of logs from the Hodrick-Prescott trend. Results
are in Table 10.3. In the prewar period the terms of trade are procyclical but the
relation is not very strong. In the interwar period the terms of trade excluding
tariffs are strongly countercyclical. The difference between the prewar and inter-
war periods was noted by Lipsey (1963, p. 15): "In the short-run behavior of U.S.
terms of trade, a sharp shift may be noted. In the prewar years ... they moved with
prices and were roughly inverse to the terms of trade of the U.K. and [industrial-
ized Europe].... After World War I, when U.S. terms of trade became inverse to
price changes, they conformed well to both British and [European] terms of trade.
It might be said that the trade pattern matured, developing from one that is char-
acteristic of a primary goods exporter to one characteristic of a nation exporting
manufactured products." In the postwar period terms of trade excluding tariffs are
weakly procyclical if 1974 (the year of the oil price shock) is included in the
sample, otherwise weakly countercyclical. The interwar pattern in terms of trade
including tariffs depends on the tariff index. Using Lerdau's index they appear
weakly countercyclical; with Crucini's indices they appear weakly procyclical.
Table 10.3 also shows the cyclical pattern in (gross) investment in capital for
household production-that is, the flow of consumer durables. Shaw (1947) presents
an annual series for this variable from 1889 through 1939. It is independent of both
the prewar and interwar industrial production indices. Deviation from trend
in durables consumption appears strongly procyclical before and after World
War I. Indeed, examination of the Shaw series shows that an absolute decline in
364 MACROECONOMETIUCS
Table 10.3. Cyclical patterns in terms of trade consumer durables investment (all
variables in logs)
1879-1914 0.0772
(0.73)
1919-1940 -0.3439 -0.1334 0.0124 0.0673
(-5.03) (-0.61) (0.0987) (0.4254)
1959-1990 0.2006
(0.86)
1959-1973 -0.1992
1975-1990 (-1.13)
1889-1914 2.9097
(91.57)
1919-1940 6.052
(81.60)
Sources: Tenns of trade 1879-1940 from Lipsey (1963, table HI); tenns of trade 1959-1990
export price deflator divided by import price deflator, U.S. Council of Economic Advisers (1993, p.
353). Tariff indices 1919-1940 from Lerdau (1957, p. 235) and Crucini (1994). Production of con-
sumer durables from Shaw (1947). Industrial production 1879-1914 is Frickey (1947) manufacturing
production index, 1919-1990 Federal Reserve Board index of industrial production.
These patterns and other research in economic history bear on many issues in
business-cycle theory. Here we focus on the debate between the two major schools
of thought in macroeconomics at present: the new classical or real business-cycle
approach versus the Keynesian approach and modern new Keynesian attempts to
fill in the holes in the old Keynesian model. We judge the plausibility of each
school's models in light of historical experience. Along the way we point out some
important open questions for empirical study.
What Causes Business Cycles? New classical models assume that cyclical-
frequency variations in output and employment are caused by shocks to real
preferences and constraints: "no matter how monetary policy is conducted, the
behavior of real quantities is determined by real shocks to the economy" (Manuelli,
1986). Such shocks might include temporary changes in the government's demand
for goods and services (as in Barro, 1981; Christiano and Eichenbaum, 1992) or
changes in the expected return to saving and investment. However, assuming con-
ventional household utility functions, either of these would cause consumption to
move in the opposite direction from output and employment. 5 Since consumption
appears decidedly procyclical in all historical periods, new classical models must
rely on another sort of real shock: procyclical fluctuations in the return to employ-
ment versus leisure (Barro and King, 1984; Mankiw, 1989). If wages reflect the
return to employment and the return to leisure is fixed, then the models imply that
real consumption wages must be procyclical. The failure of most empirical studies
to find such a pattern has been taken as evidence against new classical models
366 MACROECONOMETIUCS
(David Romer, 1993). Barro and King (1984) speculate that real wages need not
reflect the return to employment if employment relations are long-term and firms
smooth real wages to insure employees against cyclical shocks. A study of real
consumption wages over pre-19l4 business cycles could be useful here, and is a
conspicuous gap in the literature. Before World War I most workers' attachments
to firms were not long term: turnover was much higher, employment spells much
shorter, and a laid-off worker much less likely to be rehired by his old firm than
in later decades (Jacoby, 1985; Jacoby and Sharma, 1992). Prewar data on wage
rates and prices of consumption goods ought to show the underlying pattern in the
return to employment even if postwar data do not.
What might be the exogenous source of short-run fluctuations in the return to
employment? Procyclical variations in the rate at which domestic output can be
exchanged for foreign goods would do the trick (Mendoza, 1991). But they would
presumably appear as a procyclical pattern in the terms of trade and the trade
balance. As noted above, there is no strong procyclical pattern in the terms of trade
in any historical period, and the trade balance appears countercyclical in all periods.
Technology Shocks in History. One can test the new classical hypothesis about
technological change against the considerable body of research on technological
innovation and diffusion, much of which has taken place within the context of
economic history. One aspect of the cliometric revolution was the careful dating
and quantification of observations about technological changes, including estim-
ates of their effects on total factor productivity. Economic history is also useful
here because long hindsight allows one to identify the innovations that might be
big enough to make a difference in the aggregate and hence provide appropriate
tests of the new classical hypothesis.
Three innovations that have attracted the attention of economic historians,
precisely because they appear so overwhelmingly important, are steam-powered
railroads, stationary steam engines (to power machinery), and electric motors.
Railroads were introduced to the United States in the 1830s, with more than 2,000
miles of way in operation by 1840 (U.S. Bureau of the Census, 1975, series Q32l).
Robert Fogel (1964) and Albert Fishlow (1965) described the effect of railroads
on total factor productivity in terms of "social savings." They compared aggregate
output in certain years to estimates of what annual output would have been in the
absence of railroads using the next-best transportation technologies and the same
stocks of land, labor, and capital. Fogel guessed that railroad freight service alone
had increased total factor productivity by 4 to 5 percent as of 1890. Fishlow
estimated that both freight and passenger service had increased factor productivity
by 5 percent as of 1859. Either estimate implies that railroads increased total factor
productivity by a negligible amount from one year to the next, for no single year
saw a significant fraction of the rail network put in place, and there were gradual
improvements in railroad technology all along.
368 MACROECONOMETIUCS
G.N. von Tunzelmann (1978, p. 286) presents an estimate along these lines for
the effect of the steam engine on British GNP growth in the early nineteenth
century: "For Boulton and Watt engines alone (including their pirates) the social
savings over atmospheric engines can be put at about 0.11 percent of national
income in 1800. If total real output was then growing at its average rate for the
take-off years, the level of national income reached on 1 January 1801 would not
have been attained much before 1 February 1801 without James Watt. If all steam
engines, Watt and atmospheric alike, were hypothetically replaced with other
means of motive power (a combination of water and wind would be optimal), the
setback would have been about two months. These are upward-biassed figures."
There is no similar estimate for the effect of steam engines on American GNP
growth, but it could scarcely have been greater. Steam engines replaced waterpower
at a very gradual rate. It was not until the 1860s that steam produced more horse-
power (Atack, Bateman, and Weiss, 1980, p. 282).
The subsequent replacement of steam engines by electric motors was also slow.
"Steam power prevailed at the tum of the century.... By 1920, electricity had
replaced steam as the major source of motive power, and in 1929-just 45 years
after their first use in a factory-electric motors represented about 78 percent of
total capacity for driving machinery" (Devine, 1983, p. 349). Paul David (1990,
p. 359) guesses that the adoption of the electric motor accounts for "approximately
half of the 5 percentage point acceleration recorded in the aggregate TFP growth
rate of the U.S. manufacturing sector during 1919-29 (compared with 1909-19)."
That does not translate into much of an effect on total factor productivity within
a period relevant for business cycles.
The rates of diffusion for these particularly important technological innova-
tions were not unusually slow. Edwin Mansfield's (1961, p. 744) study of interwar
and postwar innovations found that "the diffusion of a new technique is generally
a rather slow process. Measuring from the date of the first successful commercial
application, it took 20 years or more for all the major firms to install centralized
traffic control, car retarders, by-product coal ovens, and continuous annealing.
Only in the case of the pallet-loading machine, tin container, and continuous
mining machine did it take 10 years or less for all the major firms to install them."
"The empirical picture is clear. Research from a wide variety of disciplines has
shown that new technologies are not diffused instantaneously into the prevailing
economic and social structure and that the pattern of diffusion can vary greatly
across technologies and industries, in some cases being drawn out over decades"
(Metcalfe, 1990, p. 17). Salter (1960, pp. 95-99) reviewed the history of the
diffusion of new technologies in many industries, in the United States and United
Kingdom, for the nineteenth and twentieth centuries. He argued that the slow dif-
fusion of technological change is largely attributable to the physical embodiment
of technological change in fixed capital, and the slow replacement of fixed capital
HISTORICAL MACROECONOMETRICS 369
regress does not appear to correspond to any event in Western economic history
since the fall of the Roman Empire. Two proponents of the new classical ap-
proach, Gary Hansen and Edward Prescott (1993), recently admitted as much.
"Although the rate at which inventions and discoveries are made may vary over
time, the stock of knowledge should not decrease. Thus, these variations are not
likely to account for the negative growth rates sometimes observed." Hansen and
Prescott suggest that "changes in the legal and regulatory system within a country
often induce negative as well as positive changes in technology" (p. 281). The
effect of institutions on economic performance is an important focus of economic
history (see, for example, North, 1990). The burst of growth in West Germany
after the monetary reform of 1948 was as large and rapid as a cyclical recovery.
But one would be hard-pressed to identify such events in American history, and
certainly one could not imagine a string of such events to account for experienced
downturns and upturns.
An emphasis on changes in regulatory policy, fiscal policy, or institutional
change as a source of "technology" shocks works against the real business-cycle
paradigm. Prior to World War I, when regulation, fiscal policy, and government
institutions were far less prevalent (Hughes, 1991), and when changes in such
policies were relatively infrequent, business cycles were more pronounced. The
post-World War II era has seen dramatic increases in government activism, and
a much greater frequency of change in fiscal and regulatory policies, while busi-
ness cycles dampened.
Keynesian Models
What Causes Business Cycles? The IS-LM model following John Maynard
Keynes (1936) and John R. Hicks (1937) is based on two assumptions: real money
balances are substitutes for capital assets, and nominal wages, prices, and debt
obligations underadjust to aggregate shocks. At least in the short run-say, from
one month to the next-they are "sticky." Keynes, Hicks, and most of their
American followers believed that wage rates were the main source of nominal
rigidities. Nominal price rigidities followed because prices were set as mark-up
372 MACROECONOMETRICS
over marginal cost. Since the 1970s textbook Keynesian models have combined
the IS-LM model with an aggregate-supply function that allows for exogenous
price shocks (supply shocks) from agriculture and foreign trade and have allowed
for nominal rigidities in product prices per se. A new Keynesian literature has
attempted to derive the features of the old Keynesian model from standard utility
functions and more fundamental constraints. Many new Keynesian models de-
scribe nominal rigidities in prices with no special reference to wage rates or even
an assumption that nominal wages adjust to clear labor markets. So far the similar
puzzle of fixed nominal debt contracts, also an important part of the old Keynesian
story, has attracted surprisingly little attention.
Purely monetary shocks, like changes in the aggregate money-demand function
or changes in the rate of growth of the money supply, have no (or minimal) real
effects in new classical models. In Keynesian models monetary shocks cause
variations in employment and real output, procyclical movements in consumption
and investment (both market and household), and countercyclical movements in
the trade balance. Monetary shocks affect real variables in three ways that operate
together but are theoretically distinct.
First, the production level chosen by a profit-maximizing firm varies when the
nominal product demand schedule shifts relative to sticky prices of products or
factors. When demand falls, for example, a firm that is constrained from cutting
the nominal wage rates it pays for given jobs will lay off employees, though the
workers' marginal revenue product remains greater than the value of their best
alternative opportunity: low-value underemployment, a search for another job,
even involuntary unemployment.
Second, whether or not wages and prices are sticky, changes in demand affect
the size of existing nominal debt obligations relative to debtors' nominal incomes.
In the presence of some plausible capital-market imperfections the real burden of
existing debt affects investment and consumption as described by Ben Bernanke
(1983). Keynes himself was confused on this point, arguing that a fall in aggregate
demand would increase debt burdens only if it were associated with a fall in wages
and prices (1936, p. 264). In fact, the real burden of a fixed nominal debt depends
on the size of the contracted payments relative to the debtor's nominal income or
wealth. Real income and wealth fall with aggregate demand even if wages and
prices are rigid.
Third, since wages and prices are not absolutely rigid, changes in nominal
aggregate demand should be associated with changes in expected inflation rates.
Changes in expected inflation affect the expected real return to holding money
balances and the demand for real assets. This implies that during the wage and
price adjustment process, a higher expected inflation rate will offset somewhat the
expansionary effect of a temporary increase in real balances produced by sticky
prices. DeLong and Summers (1986) show that for some price- and wage-setting
HISTORICAL MACROECONOMETRICS 373
rules this effect even can make business-cycle volatility an increasing function of
the degree of price flexibility because greater price flexibility increases the rate of
expected inflation during the adjustment process.
In Keynesian models, aggregate real shocks like changes in government spend-
ing or changes in the expected return to investment have effects quite different
from those predicted by new classical models. As Keynesian IS shocks, all can be
associated with procyclical movements in consumption and investment and
countercyclical movements in the trade balance. Also unlike new classical models,
Keynesian models have no necessary implication for the real return to employ-
ment. Real wages may be countercyclical, acyclical, or procyclical depending on
the relative rigidity of prices versus wages (Barro, 1979, p. 59; Mankiw, 1990, p.
1657). Real wages could be procyclical even if firms were free to choose prices
subject to sticky nominal wage rates, as long as profit-maximizing price mark-ups
over marginal cost were countercyclical (Dunlop, 1938; Keynes, 1939; Tobin,
1993). Countercyclical mark-ups can be derived from a number of different as-
sumptions about the structure of product markets (for examples, see Bils, 1989;
Rotemberg and Woodford, 1991).
The predictions of Keynesian models are consistent with the cyclical patterns in
real variables that appear in the prewar and interwar periods as well as the postwar
period. Perhaps more important, they point to specific causes of most business
cycles in terms of observable events. Narratives like those of Friedman and
Schwartz (1963) and Rendigs Fels (1959) present a long catalogue of plausible
monetary (LM) and spending (IS) shocks to account for almost every upturn and
downturn as defined by the NBER-an embarrassment of riches that has only
been carefully studied for the case of the interwar Great Depression (for example,
Temin, 1989).
The macroeconomic events of 1929-1941 pose no special problems for
Keynesian theory, unlike new classical theories. As Robert Barro (1979, p. 58) has
noted, "From a reduced-form perspective that relates business fluctuations to prior
monetary disturbances, the contraction from 1930 to 1933 seems well in line with
other experiences. The unprecedented monetary collapse over this period accords
quantitatively with the drastic decline in economic activity." Christina Romer
(1993, p. 24) argues that "the path of output and employment in the United States
in the 1930s is, contrary to common perception, very well understood" in terms
of a standard Keynesian model: "The United States slipped into recession in mid-
1929 because of tight domestic monetary policy aimed at stemming speculation
on the U.S. stock market. The Great Depression started in earnest when the stock
374 MACROECONOMETRlCS
market crash in the United States caused consumers and firms to become nervous
and therefore to stop buying irreversible durable goods. The American Depression
worsened when banking panics swept undiversified and overextended rural banks
and the Federal Reserve failed to intervene. Finally, the American Depression
ended when the Roosevelt administration chose to increase the money supply
tremendously after 1933" (p. 37). The increase in the money supply "had exactly
the effect on the U.S. economy that a traditional aggregate supply-aggregate
demand model would lead one to predict ... real interest rates plummeted in
response to the gold inflows ... followed fairly quickly by a recovery of interest-
sensitive spending, such as construction spending and consumer purchases of
durable goods" (p. 36). While Romer's description of the Great Depression may
attach too little weight to long-lived disruptions (supply-side shifts) that followed
the severe aggregate-demand shocks of 1929-1933, she is on firm ground in her
claim that the downturns and upturns of 1929-1941 are associated with the move-
ments in monetary variables predicted by the Keynesian model.
wholesale prices, and consumer prices fell along with industrial production, real
GNP, and employment (according to anyone's series).7
Fortunately, each of these downturns left behind records of changes in wage
rates paid by firms for given jobs that are quite consistent with the notion that wage
rigidity was associated with significant layoffs during business-cycle downturns in
the face of falling aggregate demand and faIling prices. The NBER dates the 1893
depression as a peak in January 1893 and a trough in June 1894. In 1894 the state
labor bureaus of New York, Ohio, Connecticut, and Maine surveyed establish-
ments as to whether they had cut wage rates over the calendar year 1893 or in a
period from early 1893 through summer 1894. The Ohio bureau continued its
surveys through the 19OOs, covering the NBER peak in May 1907 and trough in
June 1908. These state surveys are described by William Sundstrom (1990, 1992),
Carter and Sutch (1990), and Hanes (1993). In every survey, many establishments
reported cuts in wage rates for some or all occupations. This is consistent with
available time series on average wage rates or hourly earnings, which show de-
creases over the same years. But in all but one survey, most establishments re-
ported they had not cut wage rates for any occupation, even though they had cut
employment by large percentages. The Ohio 1894 survey did not ask about em-
ployment, but other data show that employment fell a lot in the same sectors and
geographic area (Sundstrom, 1990). The exception is the Connecticut 1894 sur-
vey, which includes the longest period after the downturn (through August 1894).
In that survey slightly more than half of the establishments reported wage cuts
(Carter and Sutch, 1990, p. 17). In the early 1930s the V.S. Bureau of Labor
Statistics collected similar (but not directly comparable) information on wage rate
changes. These data show that many establishments held wage rates fixed until
more than a year after the downturn from the NBER peak in August 1929, long
after they had cut employment (Shister, 1944; Mitchell, 1985).
There is no way to be sure that the behavior of the firms in these samples was
representative of firms in general, but the surveys' results are certainly consistent
with other information about firms' wage rate changes in these downturns. Robert
Ozanne's (1968) study ofInternational Harvester, originally the McCormick Reaper
Company, shows that the firm cut employment early in each of the three depres-
sions. But in 1893 the company "cut only the [wages of] skilled workers and those
on piece rates. Eventually, three years after the depression began, the McCormick
firm cut [wages for] common labor I cent" (Ozanne, p. 32). After the 1907 down-
turn International Harvester did not cut wages until the beginning of 1908 (p. 37).
After 1929 "the company took no action to cut wages.... Finally, in October,
1931, the first production worker wage cuts of 15 percent were instituted" (p. 52).
In 1908 V.S. Steel and other large steel producers refrained from wage cuts despite
deep reductions in employment (Shergold, 1975, p. 181). After 1929 V.S. Steel
held its common labor wage fixed until October 1931. Other firms appear to have
376 MACROECONOMET~CS
followed the same policy. "If we believe the newspapers, magazines, trade pub-
lications, and academic journals of the time, money wages in manufacturing were
not cut until the fall of 1931" (O'Brien, 1989, pp. 720, 721; see also Jacoby, 1985,
pp. 216, 217). There can be no reasonable doubt that in each of these downturns,
months after wholesale and consumer prices had begun to fall, establishments
accounting for a significant fraction of aggregate employment and output held
their nominal wages absolutely fixed as they cut production and laid off workers. s
Note that all of this evidence is in regard to wages in just one sector-
manufacturing. It is important to note that nominal wage rigidity in the manufac-
turing sector need not have had the same effects on aggregate output and employ-
ment in the prewar and interwar periods as in the postwar period. As we have
already mentioned, agriculture is an acyclical sector, and it was much more im-
portant before World War II. Unlike manufacturing wages, farm wages appear to
have adjusted quickly to aggregate shocks (Goldin and Margo, 1992), and at least
in the nineteenth century, workers moved between the manufacturing and farm
sectors at seasonal and cyclical frequencies (Dawley, 1976; Keyssar, 1986; Goldin
and Engerman, 1993).
The decline in the intersectoral substitution of labor is also related to the de-
cline in labor turnover within manufacturing over time. Margo (l990a, 1990b,
1991) shows that the duration of unemployment spells in the manufacturing sector
was relatively short prior to the Great Depression, and that long-duration unem-
ployment became much more common during and after the 1930s. Some of the
high historical turnover rate (short unemployment spell duration) in manufactur-
ing likely reflected substitution into agriculture during cyclical and seasonal lulls
in manufacturing.9 The elasticity of labor substitution across sectors, and the turn-
over of labor, was reduced in the early twentieth century (Jacoby and Sharma,
1992).
The relative ease with which labor moved across sectors before World War I,
combined with the smaller size of the manufacturing sector relative to farming, the
acyclicality of agricultural production, and the more rapid adjustment of nominal
wages in agriculture than manufacturing all made it easier for displaced manufac-
turing workers to be absorbed in agriculture. Thus, a cyclical downturn in manu-
facturing and other sticky-wage sectors could have pushed workers into agricultural
underemployment rather than unemployment.
Wage Rigidity and the Phillips Curve. In our discussion of wage rigidity, we
have so far avoided reference to results of the Phillips curve studies referred to
earlier. That is because none of those studies makes much effort to relate the
coefficients of estimated relations to an underlying model of the system of setting
prices, wages, and output. Under any reasonable assumptions, the estimated rela-
tions are at best reduced forms of much more complex systems, and interpretation
HISTORICAL MACROECONOMETRICS 377
Menu Costs. The most general explanation of nominal wage and price rigidity
offered by new Keynesian models is the notion of menu costs. Menu cost models
show that a firm's incentive to adjust prices to shifts in nominal aggregate demand
can be small if its product market is less than perfectly competitive (Mankiw,
1985). As Robert Gordon (1990) and David Romer (1993) point out, the menu-
cost argument also applies to wages, whether they are set by firms or monopoly
unions. The less elastic is product demand, the smaller the increase in profit result-
ing from a decrease in marginal cost and hence the smaller the firm's incentive to
cut wages. The firm's derived demand for labor is also less elastic, so allowing
a wage cut gives workers a relatively small increase in employment (a small
decrease in the probability of being laid off).
If the benefit of adjusting prices or wages is small, a small "menu" cost of
adjustment itself is enough to cause the setter to hold them fixed. It is not clear,
however, what these menu costs might be. So far proponents of these models have
invoked various forms of transactions and administrative costs. "Taken literally,
these menu costs are the resources required to post new price lists. More
378 MACROECONOMETRICS
metaphorically, these menu costs include the time taken to inform customers, the
customer annoyance caused by price changes, and the effort required even to think
about a price change" (Mankiw, 1990, p. 1657). But failing to adjust prices brings
on administrative costs of adjusting output and employment-costs of cutting
production, laying off workers (with the associated turnover costs), closing down
plants altogether-which seem larger, if anything.
What kind of costs could apply to wage and price changes alone? The Great
Depression offers an example: public disapproval, government pressure and regu-
lations. Anthony O'Brien (1989, p. 724) argues that in the 1920s it was generally
believed that wage and price deflation made depressions worse, or even caused
them: "By the mid-1920s many manufacturing firms had publicly announced that
wage rates would not be reduced during the next downturn." Shortly after the
stock-market crash President Hoover called a conference of large employers to
encourage them to maintain nominal wage rates. Most promised to do so. Thus
"cutting wage rates at the onset of the depression" would have been a "public
relations debacle" (p. 729). In 1933 the National Industrial Recovery Act created
minimum wages and other forms of regulation that raised nominal wages and
prices after deflation had set in and before any considerable upturn in employment.
Michael Weinstein (1981, p. 279) points out that "the economy could not have
recovered in historically expected ways as long as the NIRA wage and price
regulations were effective."
These policies were peculiar to the 1930s, of course, but it may be useful to
consider the reasons that rational government and business administrators came
to adopt such policies, and whether they have general implications. Olmstead and
Rhode (1985) argue that similar "social pressures" were important in restricting
gasoline price increases during the California gasoline famine of 1920. Gasoline
sellers voluntarily chose to ration fuel rather than increase its price to the market
clearing level, which Olmstead and Rhode argue reflected their fears of violent
public reactions to price increases.
implicit state-contingent contract that set employment at the optimal level, where
the marginal revenue product of labor equaled its opportunity cost and stipulated
payments to workers so as to equalize the marginal utility of income across states.
Under such an agreement the nominal or real "wage rate" prevailing at any point
in time would be nothing more than the state-contingent payment divided by the
state-contingent quantity of labor supplied. It would not be a price in the ordinary
sense and would have no direct relation to employment (Barro and King, 1984).
This argument seems to have been accepted by some new Keynesians, contrib-
uting to the recent emphasis on price rather than wage rigidity, though it applies
just as well to product prices: firms can also have long-term relationships with
suppliers of raw materials and with consumers of final products. Fortunately, later
work following Calvo and Phelps (1977) showed that the neutrality argument
depends on an implausible assumption that workers know as much as firms about
the value of labor's product. If a firm has private information and is willing to
trade off some amount of expected profit for a reduction in its variance-a con-
sequence of the same financial-market imperfections that generate real effects of
a debt burden-then the optimal contract leaves the firm free to choose the em-
ployment level subject to positive marginal costs of labor (Azariadis, 1983), just
as in the old Keynesian story. Thus, the problem becomes one of accounting for
the underadjustment of marginal labor costs-that is, wage rates (and employ-
ment-contingent benefits)-to aggregate shocks.
Many Keynesian models of the late 1970s accounted for the underadjustment
of nominal wage rates with an assumption that wages were set by long-term
contracts before the realization of aggregate shocks. This was often justified as a
depiction of American union labor contracts, which set nominal wage rates for one
to three years with little or no indexing. The relation between union contracts and
nominal wage rigidity seemed to be confirmed by the earliest studies of wage and
price behavior based on standard historical series for output, prices, and average
hourly earnings. As we mentioned above, Cagan (1975) and Sachs (1980) showed
changes in Phillips curve coefficients on output and lagged inflation from prewar
or interwar periods to the postwar period. That matched up nicely to the timing of
the spread of unions and long-term wage contracts across the manufacturing sector.
Robert Gordon (1986, p. 20) summarized the conventional wisdom: "The greater
postwar persistence of wages and prices is generally attributed to two factors.
First, the increased importance of labor unions since the late 1930s has led to
centralized wage bargaining, and high perceived costs of negotiation have made
it economical to establish three-year contracts in many industries. That today's
wage changes were in many cases agreed upon last year or the year before tends
to insulate wage changes from current market forces and to increase their depend-
ence on what has happened previously. Second, the greater confidence of private
agents in the willingness of monetary and fiscal policy to reduce the severity of
380 MACROECONOMETRlCS
recessions lessens their need to reduce wages and prices quickly and increases
their incentive to wait for the expected prompt return of prosperity."
But pointing to union contracts or any other long-tenn price arrangements did
not amount to an explanation of nominal rigidity. Why should workers and firms
fail to adjust the marginal cost of labor, determined by the tenns of the contract,
to aggregate shocks that both sides should be able to observe? Wage bargainers
should adjust rates set by a long-term contract to those shocks, either by renego-
tiating the contract or indexing its tenns in advance. As Robert Barro (1979, p. 54)
put it, "Long-term labor agreements do not imply a failure of employment to
increase when all parties to the agreements perceive that they could be made better
off by such a change." Perfect indexation of a long-tenn contract may be impos-
sible, if only because published price indices are very noisy measures of true
changes in the price level (Meltzer, 1994; Bullard, 1994). But it is hard to account
for a total lack of indexation in a three-year contract or a failure to renegotiate a
contract in the face of a general recession.
More important, wage rates display nominal rigidity in the absence of any kind
of long-term contract. The historical association between union contracts and
wage rigidity has not held up to further scrutiny. Recall that better wage and price
series used in more recent studies of the Phillips curve show no decrease in the
Phillips curve coefficient on output (Allen, 1992; Hanes, 1993). The coefficient on
lagged inflation still appears larger in the postwar period, but as we have argued,
that did not reflect a change in the association between wages or prices and de-
viations of output from trend. Most important, the microeconomic historical evid-
ence we have discussed indicates that many firms refrained from cutting wages in
prewar and interwar depressions like 1893, 1907, and 1929 though few faced
fonnal unions and none was constrained by a contract. Before the 1930s courts did
not enforce wage contracts between finns and worker groups even if they were
written down.
Many patterns in postwar data are consistent with the ideas in the insider-
outsider literature. Some nonunion firms appear to pay wage premia that resemble
those in unionized firms-in excess of compensating differentials for unpleasant
work conditions or differentials associated with an employee's human-capital
characteristics and paid to workers in virtually all the firm's jobs. High-wage firms
are usually found in industries with certain characteristics: large establishments,
firms with high capital-labor ratios, and firms that enjoy greater product market
power (those with high profits and market shares). For example, wages are higher
in manufacturing than in agriculture and higher in machine tool manufacture than
in textiles. Establishment size, capital-intensity, and market-power indicators are
highly correlated across sectors, so it is hard to identify which might bear a causal
relation to wage premia (Dickens and Katz, 1987), but the latter two correlations
are certainly "consistent with the idea that labor can extract rents which depend
on how much damage it could do by temporarily stopping the firm from produc-
ing. Production interruptions are more costly for capital-intensive firms and those
earning high profits than for other firms" (Katz and Summers, 1989, p. 241). Post-
war data also suggest an association between high wages and nominal wage rigidity.
Wage differentials across sectors and firms are countercyclical: wages in high-
wage sectors fall less in recessions and rise less in booms (Wachter, 1970). But
since firms in high-wage sectors are also more likely to be unionized, it is hard to
say whether this pattern holds for high-wage nonunion firms.
Historical evidence allows one to separate the effects of workers' bargaining
power and industry characteristics from those of formal unions and long-term
contracts, since unions were rare before the late 1930s, even in manufacturing.
Such evidence makes it clear that worker groups can threaten firms even in the
absence of formal organization. In the late nineteenth century nonunion workers
were capable of withdrawing effort in the most spectacular way, by going on strike
(Hanes, 1993). Large-scale strikes by nonunion workers against cuts in nominal
wages and threatened cuts continued through the 1920s (Wright, 1981), and it was
observed that nonunion workers would slack off and damage machinery if they
felt they had been treated "unfairly" (Mathewson, 1931). The firm characteristics
that are associated with wage premia and unions in the postwar period were as-
sociated with strikes in the late nineteenth century (Hanes, 1993) and wage premia
in the 1900s and 1920s, before the spread of unions (Krueger and Summers,
1987).
They were also associated with nominal wage rigidity. In the downturn after
1929 firms in industries with less competitive product markets and high capital-
labor ratios were the last to cut wages (Shister, 1944). In the depression of 1893
firms were less likely to cut wages if their industries had suffered especially large
numbers of strikes in earlier years (Hanes, 1993). Recall that the general historical
pattern in the cyclica! behavior of nominal wages and prices appears (at present)
382 MACROECONOMETRICS
Historical evidence suggests that strategic threats and bargaining have been
more important. Daniel Raff (1988, p. 398) has examined Henry Ford's adoption
of the "five-dollar day" in 1914, a clear example of a policy to pay wages in excess
of workers' alternative opportunities: "None of the three canonical efficiency
wage theories-turnover costs, adverse selection, and moral hazard-speaks to a
plausible central motive. A desire to buy the peace seems far more consistent with
the facts. Ford was alone in his industry in paying such wages when he did because
he was alone in his method of making automobiles." Ford faced "the specter of
the sit-down strike. Ford's single-minded focus on a dedicated capital stock and
economies of scale made the company peculiarly vulnerable. Its profits left the
opportunity cost of such action to the owners of the company relatively high....
The Wobblies had been in Detroit in the late springtime of 1913" (p. 396).
Sanford Jacoby (1985) describes the origins of the internal labor-market poli-
cies that should be related to the payment of premium wages, if efficiency-wage
models are correct. Jacoby finds that "the new bureaucratic approach to employ-
ment did not gradually take hold in an ever-growing number of firms" (p. 8), as
one might expect if it were a profit-maximizing managerial innovation. Instead,
"the growth of internal labor markets in large firms was erratic, occurring chiefly
during World War I and the Great Depression" (p. 280). At the same time firms
came to see turnover as a bad thing, but "the new concern with turnover was not
attributable to any increase in separation costs associated with technological change
of a firm-specific character. In fact, manufacturing technology was becoming less
firm-specific and idiosyncratic than it had been in the nineteenth century" (p. 122).
"The historical record indicates that the employment reforms introduced during
World War I and after 1933 were attributable not so much to competitive market
forces as to the growing power of the unions and the ascendance of the personnel
department over other branches of management. ... With the benefit of forty
years' hindsight, we may observe that these policies often enhanced efficiency by
reducing turnover and increasing morale, or by stimulating programs to upgrade
the work force. But this effect was by no means obvious to the managers of firms
in transition, who were skeptical that internal labor market arrangements would
lower costs. Efficiency incentives were neither strong enough nor obvious enough
to produce the modem internal labor market" (pp. 277, 278).
Moreover, "personnel managers can be viewed as mere conduits for union
threat effects in that many of their ideas were borrowed from the unions and their
influence often rested on the imminence of labor unrest" (p. 281). The firms that
adopted internal labor-market policies in the 1900s did so to "divert industry's
skilled workers from the unions by means of quasi-pecuniary incentives, including
profit-sharing, pension, and home ownership plans" (p. 52). When unions spread
during World War I: "Personnel management reduced the potential for unioniza-
tion by replicating within the firm some of the union's protective structures"
384 MACROECONOMETRICS
(p. 164). After unions were crushed in the early 1920s personnel departments lost
influence, but in the 1930s, after the NIRA and the Wagner Act: "The strategy of
using employment reforms to deter unionism was given new life" (p. 242). Per-
sonnel managers cared about turnover because "the same conditions that gave rise
to quits and low morale in the present might lead to strikes or worse in the future.
If these conditions could be identified and remedied, they reasoned, morale would
improve, quits would fall, and the possibility of industrial revolt would be greatly
lessened" (p. 120).
from the number of plausible theoretical models that might explain long swings.
Similar patterns of phenomena within each cycle suggest a common mechanism.
The phases of the Kuznets cycle are remarkably common across the various cycles
and are associated with similar sequences of activities-most notably, the con-
struction of transportation networks and structures and massive waves of immigra-
tion. These similarities led proponents of the Kuznets cycle to describe long swings
as building cycles (see Abramovitz, 1964; the review of this vast literature in
Hughes, 1990, ch. 8 and 16; Solomou, 1987; Zamowitz, 1992).
There is widespread disagreement about how such cycles began-with possi-
ble triggers ranging from immigration shocks to financial or political factors that
set the stage for expansion. lO There is similar disagreement about the factors that
brought such cycles to an end. I I Whatever the answers to those questions, it is
clear that land settlement, canal building, and railroad speculation on the frontier
were central to the booms of the nineteenth century, and reversals in these pro-
cesses were associated with rapid declines in migration westward and investment.
For example, migration westward in April of 1857 resulted in the arrival of 1,000
people per day into Kansas alone. Railroads-which were speculating heavily in
land and building new lines in anticipation of a transcontinental trunk line-
encouraged the migration by lowering the cost of passage by as much as 25 per-
cent. Land prices soared, mortgage lending grew, and speculative railroad stocks
boomed. Adverse political news in the middle of 1857 (associated with the conflict
over slavery in the territories, and its consequences for timely construction of a
transcontinental line) seems to have been responsible for a collapse of speculative
land and railroad securities prices, and the end of rapid immigration (Calomiris
and Schweikart, 1991). This experience illustrates the importance of expectations
for migration decisions and land and railroad construction, and the potential for
bootstrapping effects (mutually reinforcing expectations and investments). As
Brinley Thomas (1954) notes, supply-side and demand-side forces often rein-
forced each other in producing economic growth, as international and interregional
migration responded to differences in standards of living across locations. Analy-
sis of the timing and location of persistent bursts of economic growth also show
that long swings were closely tied to specific locations in which economic devel-
opment was occurring very rapidly.
The common spatial patterns of long-run cycles historically in the United
States suggest that a model of endogenous growth capable of explaining the Kuznets
cycles must begin with a spatial conceptual framework for modeling economic
activity (e.g., Murphy, Shleifer, and Vishny, 1989; Krugman, 1991). Waves of
American economic development were associated with three related spatial
phenomena-the settlement of an expanding frontier, the establishment of a trans-
portation and communication network within the boundaries of that frontier, and
HISTORICAL MACROECONOMETRICS 387
the creation of high concentrations of population and economic activity that served
as nodes within that network. Epochs of great expansion were times when large
investments in transportation and communication occurred, removing preexisting
barriers to the flow of capital and labor and allowing the reorganization of the
production and distribution process within the economy (Chandler, 1977). These
investments promoted massive flows of capital and labor, which chose to move to
new locations and often to concentrate in large cities.
This spatial interpretation of the Kuznets cycle suggests four related features
of the spatial economy that seem essential to explaining moments of sudden
change followed by persistent expansion or contraction. First, the decision to bear
fixed costs is central to the suddenness of economic expansion that characterizes
Kuznets cycles. Entrepreneurs must be willing to pay fixed costs necessary to the
establishment of transportation and communication facilities that substantially
reduce the marginal cost of settlement. Once the initial barriers are breached, that
sets the stage for a prolonged expansion-a free lunch for followers and, of course,
some rents for leaders.
Second, the expectations of entrepreneurs are central to the beginning of this
process. That does not mean that visionary entrepreneurs "make it all happen" (for
example, see Fishlow's, 1965, analysis of patterns of settlement and railroad con-
struction in the antebellum period). But it does mean that rapid settlement must
be occurring, or expected to occur soon, to justify large expenditures of resources
on infrastructure. Moments of expansion are focal points when immigrants, land
speculators, infrastructure builders, and external owners of capital agree that rents
can be had from moving quickly into new areas to build and populate them or to
connect them to preexisting areas in new ways. The frontier always beckoned
as the obvious point of future expansion, but the timing of the great push was
unknown. Clearly, the origins of long waves cannot be attributed to exogenous
technological changes. The technologies for building canals and railroads or for
harvesting the vast prairies of the Midwest were developed long in advance of
their application. Demand-side explanations for origins (including financial and
political events, domestically or abroad) are more plausible as focal points for
starting or stopping long waves, but it would be hard to argue that uniquely large
shocks started long waves. Convergent expectations-often associated with
dramatic aggregate-demand shocks (like wars)-seem to be a key ingredient to
explaining why some shocks helped to trigger great pushes of economic activity
or great collapses. Of course, the precise timing of convergent (perhaps mutually
fulfilling) expectations is intrinsically hard to explain convincingly.
Third, in addition to the removal of fixed costs of settlement, persistence of
growth seems to be related to economies of agglomeration (city building). Location-
specific factor market and product market externalities are the defining characteristic
388 MACROECONOMETIDCS
Econometric Lessons
What are the methodological implications of historical long swings for macro-
econometrics? First, as already noted, demand shocks often set in motion long-run
changes in aggregate supply, and thus it is wrong to view long-run shifts in supply
as independent of momentary demand disturbances. Second, the importance of
spatial concepts like cities and frontiers suggest advantages from disaggregation
across locations for understanding macroeconomic change. Third, the number of
observations of business cycles may be few, even though the number of calendar
years of data are many. Once one defines the cycle of interest to be of eighteen-
year duration, that substantiaIly limits the number of cycles. This suggests further
advantages from disaggregation. By relying on panel data (for states and towns,
enterprises and individuals), one can exploit cross-sectional variation in the data
when testing hypotheses. As we show in the subsequent discussion, recognition of
institutional nonneutrality reinforces these same methodological implications.
Institutional Nonneutrality
Labor Markets
In addition to basic legal and ethical principles on which a market economy must
be founded, other institutions play an active role in determining outcomes. In labor
contracting, for example (discussed at length above), some macroeconomists and
390 MACROECONOMETRICS
historians have placed great weight on technological changes and the development
of unions in affecting the bargaining power of workers, the distribution of wealth,
and the rigidity of wages. Changes in the integration of labor markets have also
been emphasized as important preconditions to the settlement of the frontier and
the transfer of workers from relatively unproductive occupations to more productive
regions and jobs (Field, 1978; Wright, 1979, 1981, 1986; Hatton and Williamson,
1991; Goldin and Sokoloff, 1984; Rosenbloom, 1988, 1990; Sundstrom and
Rosenbloom, 1993; Margo, 1994; Sirnkovich, 1994). Such changes were impor-
tant determinants of the cross-regional distribution of wealth and were necessary
conditions for rapid expansion phases of the Kuznets cycle.
widened in the latter half of the nineteenth century as the geographic scope of
the economy and the minimum efficient scale of production expanded rapidly
(Chandler, 1977; Calomiris, 1993b; Bodenhorn, 1992; Atack and Bateman, 1994).
Small unit banks were unable to meet the needs of the new generation of large
firms operating production and distribution networks throughout the country. The
participation of banks in industrial finance was relatively small, as banks came to
specialize more in financing commercial enterprises. The costs of industrial finance
in the United States were much higher than in the concentrated, universal banking
system of Germany, where large-scale banking permitted greater diversification,
as well as the reaping of network and scope economies in placing traded securities,
managing trust accounts, and participating in corporate finance and governance of
large-scale industrial firms (Calomiris, 1994):2 Higher financing costs not only
retarded industrial expansion, but also led industrial producers to favor less fixed-
capital intensive production processes (Field, 1983, 1987; Wright, 1990; Calomiris,
1994).
Access to bank services in rural areas was also restricted by unit banking laws.
Unit banking reduced the profitability of establishing high-overhead banks in
sparsely populated areas, which would have been better served by low-overhead
branches (Evanoff, 1988). Furthermore, after episodes of local or regional finan-
cial distress, states with unit banking laws found it much harder to rebuild their
banking systems through acquisitions of failed banks or the establishment of new
offices (Calomiris, 1990, 1992, 1993b).
While the mortgage market-which was able to rely on insurance companies
for financing-apparently overcame some of the barriers to integration posed by
unit banking laws, Snowden (1993) and Snowden and Abu-Saba (1994) argue that
the fragmentation of the national mortgage market (a by-product of unit banking
laws) exacerbated agency problems between mortgage financiers (insurance com-
panies) and local brokers and promoted inefficient allocation of mortgage credit.
The greater instability of the unit banking system-which saw much higher
failure rates and a unique propensity for banking panics in comparison with other
countries-was also a result of the fragmentation of the financial system that unit
banking entailed. Times of trouble for banks both reflected adverse economic
conditions and propagated cyclical disturbances through contractions in the
availability of credit (Calomiris and Hubbard, 1989; Calomiris and Gorton, 1991;
Grossman, 1993). Other countries enjoyed the benefits of the diversification of
risk ex ante, and the coordination of the banking system's response to adversity
ex post, which a banking system with a small number of large banks made possible
(Calomiris and Gorton, 1991; Calomiris and Schweikart, 1991; Calomiris, 1993b).
In addition, unit banking supporters eager to achieve banking stability without
allowing branching pushed for the establishment of deposit insurance, which
further undermined the stability of the system by encouraging excessive risk
392 MACROECONOMETRICS
taking by banks (White, 1983; Calomiris, 1993b; Calomiris and White, 1994;
Wheelock and Kumbhakar, 1994).
Despite the potential benefits of relaxing unit banking restrictions, special in-
terests succeeded in blocking branch banking legislation (White, 1984; Calomiris,
1993b; Calomiris and White, 1994). Historical "path dependence" played an im-
portant role here. Prior to the I870s, branch banking was not a contentious issue,
and the costs of unit banking were arguably small, since the minimum efficient
scale of industrial firms and the geographic scope of the economy were limited up
to that point. During the second industrial revolution, as unit banking increasingly
became a costly constraint on the economy, the existing special interests of unit
bankers and their allies limited the adaptation of the banking system to the new
needs of industry, agriculture, and commerce.
The consequences of unit banking laws for the financial system helped set
the stage for the creation of the Federal Reserve in 1913. Like most changes in
government financial regulation, the Fed was created in the wake of a crisis-the
Panic of 1907-as a brainchild of the National Monetary Commission, which
brought together politicians, bankers, and many of the most prominent financial
economists of the time to redesign the American banking system. Given the poli-
tics of unit banking, however, the Federal Reserve Act avoided dealing with the
fragmentation of the banking system, which was understood to be the fundamental
source of instability in the system. Instead, the Fed was designed to stabilize the
banking system by limiting interbank lending and reducing the seasonal volatility
of interest rates.
The Fed was conceived essentially as a microeconomic intervention into the
reserve market. It was intended to remedy, first and foremost, the absence of an
elastic supply of reserves at seasonal frequency. Given seasonal fluctuations in
loan demand-linked to the needs to finance planting, harvesting, and marketing
of agricultural produce-banks in the aggregate were forced to choose between
allowing seasonal fluctuations in their loan-to-reserve ratio or importing and
exporting gold in large quantities in response to seasonal fluctuations in loan
demand. 13 Gold was costly to import and export (it cost roughly 0.5 percent of the
value of gold to bring it across the Atlantic) (Officer, 1986). Thus there were
strong incentives to allow loan-to-reserve ratios to fluctuate with demand. Sea-
sonal increases in the loan ratio placed banks at greater risk of defaulting on their
deposits and therefore resulted in the comovement of loan interest rates and loan
demand. Thus peak-time borrowers were forced to compensate bankers for the
greater portfolio risk to banks during peak periods. Clark (1986), Miron (1986),
Barsky et al. (1988), Calomiris and Gorton (1991), and Calomiris and Hubbard
(1994b) discuss evidence that prior to the founding of the Fed seasonal fluctua-
tions in U.S. loan demand produced important seasonal fluctuations in U.S. and
foreign interest rates, and in flows of foreign capital and gold. The Fed's discount
HISTORICAL MACROECONOMETRICS 393
by the constraints posed by special interests. The Fed is also a good example of
Field's (1981) argument that institutions come to take on a life of their own inde-
pendent of the reasons they came into being. Countercyclical monetary policy,
which came to be the Fed's primary occupation, was not one of the principal
objectives that gave rise to the Fed.
One way to approach the economic history of the period leading up to and follow-
ing the Great Depression is to consider it as a case study in institutional adaptation
to changing circumstances. The emphasis of much of the literature on the period
1929-1939 is that learning was occurring-about appropriate monetary policy
rules, about the macroeconomic consequences of wage rigidity, about the costs of
adherence to the gold standard, and about the remaining weaknesses in the bank-
ing system. Interestingly, the recent literature on institutional learning during the
Depression does not suggest that, as a general proposition, learning led to im-
provement in policy. Rather, the record is mixed. In some respects, the right
lessons seem to have been learned-particularly with respect to monetary policy
rules, the disadvantages of rigid wages, and the potential costs of adherence to the
gold standard. In other respects-notably, bank regulatory policy and fiscal policy-
recent research suggests that major institutional changes resulting from govern-
ment intervention were ill-advised and based on false interpretations of the
Depression.
On the positive side, the Fed's approach to monetary policy seems to have
changed for the better. The increased concentration of Fed authority in Washing-
ton, and greater coordination of decision making within the Fed, which Marriner
Eccles championed through his support for the Banking Act of 1935 and his
stewardship as Fed chairman, transformed the Fed into a more effective (though
not perfect) maker of monetary policy. Friedman and Schwartz (1963) provided
the seminal analysis of policy errors by the Federal Reserve during the Depres-
sion, which has been substantially updated by recent research (see the review in
Calomiris, 1993c). Friedman and Schwartz argued that Fed money-supply con-
tractions produced the Great Depression and that those errors would not have
occurred if Benjamin Strong had remained at the helm of the New York Fed. The
thrust of much of the recent work on Fed policy during the interwar period casts
doubt on these arguments. Despite strong evidence that monetary policy was
contractionary in 1929 (which was not the focus of Friedman and Schwartz's
argument), later episodes of monetary contraction seem likely not to have resulted
from money-supply shocks. Moreover, Fed policy was constrained by adherence
HISTORICAL MACROECONOMETRICS 395
to the gold standard and by concerns over gold outflows, which limited its reac-
tions to adverse shocks originating elsewhere in the economy (Eichengreen, 1992,
pp. 117-119,296-298). With respect to alleged changes in Fed targeting rules,
Wheelock (1989, 1990) has shown that the reaction function of the Fed was
essentially unchanged before and after the departure of Strong. This reaction
function was based on targeting rules of thumb that had proved useful in the past
but that misled the Fed during the Depression. It may be that the Fed's reaction
function and its adherence to the gold standard were disastrous during the Depres-
sion; however, as Temin (1989) argues, it is hard to blame the Fed for not learning
the lessons of the Great Depression before it happened.
The abandonment of the gold standard was another important component of
Great Depression learning in many countries. Adherence to the gold standard was
one of the central tenets of classical economic policy (Temin, 1989). The classical
approach-whose influence over central banking policy can be traced at least as
far as the Peel Act of 1844 in England (Helms, 1939; Wood, 1939)--emphasized
the benefits of rules over discretion. A related classical tenet was that recessions
were necessary means of forcing resources out of relatively unproductive, older
firms, and into new technologies-recessions were the occasional growing pains
of capitalism. While these tenets may have merit, slavish adherence to simplistic
rules, or the belief that recessions are always good for the economy, can have
drastic adverse consequences: this was an important lesson of the Great Depres-
sion. Countries that abandoned gold in 1931 to halt the free fall in their economies
avoided the worst of the collapse of the Great Depression, while those that re-
tained their ties to gold suffered worsening deflation, bankruptcy, and economic
disaster (Eichengreen and Sachs, 1985, 1986; Temin, 1989; Bernanke and James,
1991; Eichengreen, 1992; Bernanke, 1994). Ironically, in some cases (notably
Germany) earlier experience with hyperinflation had led government to establish
new mechanisms in the 1920s to prevent discretionary relaxation of the gold
standard (Eichengreen, 1992, pp. 125-152, 273-277). This illustrates the diffi-
culty of applying lessons from history. The shocks of the future are not always the
same as those of the past.
Eichengreen (1992) argues that the lesson of the interwar collapse under the
gold standard was not that the gold standard per se is a bad system but, rather, that
the successful maintenance of the gold standard requires coordination among
countries to prevent destabilizing exchange market disequilibrium. According to
this view, the classical (pre-World-War-I) gold standard operated well largely
because of coordination among the major central banks of Europe-in Britain,
France, and Germany-who acted together to reinforce each other's actions. This
lent credibility to anyone country's policies and discouraged private capital from
taking positions opposite central banks (that is, engaging in speculative attacks on
a currency). This may have been an important contributor to the ease of international
396 MACROECONOMETIDCS
capital flows and the small differences in money-market interest rates across coun-
tries (Calomiris and Hubbard, I994b).
While the postwar Bretton Woods system claimed to offer the possibility of
both adherence to rules and (contrary to the gold standard) the flexibility of con-
trolling the overall supply of the unit of account (that is, the dollar), it did not
provide effective checks and balances to coordinate monetary policy among the
participants. The lack of discipline of U.S. monetary policy, in particular, turned
out to be an important omission, which led to the demise of the system (Bordo and
Eichengreen, 1993). In that sense, the lessons of the collapse of the gold standard
were not learned: in retrospect, the classical gold standard of the pre-World-War-
I era remains the most successful example of how flexibility in response to crises
can be combined with long-run adherence to rules in an international monetary
system (Bordo and Kydland, 1990; Eichengreen, 1992). In this respect, history has
been regressive.
One example of progress in economic thinking and economic policy after the
1930s-which may have reflected the experience of the Depression-is the
demise of the doctrine of high wages, which seems to have no adherents currently.
According to this doctrine, resisting wage cuts during recessions is a desirable
macroeconomic policy because it keeps workers' consumption high (O'Brien,
1989). This argument is logically flawed: while it may be beneficial for individual
firms to maintain high wages in the face of a declining demand for labor (as in
some new Keynesian arguments about efficiency wages), in the aggregate this will
have a negative effect on output, not the positive effect claimed by supporters of
the doctrine of high wages. Nevertheless, this doctrine was widely believed as late
as the 1930s.
The Depression may have illustrated problems with some existing policy rules
and prescriptions, but it also provided fertile political ground for flawed economic
analyses and policies. Two prominent examples include the many changes in the
regulation of banks and the taxation of corporations. In the latter case, the surtax
on undistributed profits of 1936 is a prime example. BerIe and Means (1932) sug-
gested that lack of discipline over corporate management could lead to wasteful
use of shareholders resources. Some proponents of the classical view of business
cycles carried that argument further and argued that declining industries had con-
tributed to the Depression by refusing to payout their earnings to stockholders,
which limited resources available to finance new, more productive enterprises.
The proposed solution was the surtax on undistributed profits, which taxed firms
up to a marginal surtax rate of 27 percent on retained earnings (defined as net
earnings less dividends, calculated after the payment of normal corporate taxes).
Corporations opposed the tax from the beginning. Ironically, the most bitter op-
ponents of the tax were corporations in high-growth sectors, which tended to face
high costs of raising external funds. For these firms, taxation of retained earnings
HISTORICAL MACROECONOMETRICS 397
imposed a large cost on their primary means for financing growth (Calomiris and
Hubbard, 1994a). Compared to the presumed gains from enhancing corporate dis-
cipline, these costs were enormous. Fully 20 percent of taxpaying firms-typically
small, growing firms in high-growth industries~hose to pay marginal surtax
rates of at least 22 percent to be able to retain earnings. The opposition to the
surtax effectively won repeal of the measure after only two years.
The attack on the banking system, based on what proponents claimed were
lessons from the Depression, was much more far-reaching and long-lived. The
Banking Act of 1933 fundamentally changed the structure of financial institutions,
and its key provisions remained essentially unchanged for six decades. 16 Many of
the changes in 1933 were policies that had long been contemplated and rejected
by Congress (like federal deposit insurance). Others were variations on old themes.
For example, the separation between commercial and investment banking, and
Regulation Q, were the brainchildren of Senator Carter Glass. Glass echoed de-
cades of previous arguments about bank instability when he argued that banks'
involvement in the securities market and the pyramiding of reserves had promoted
the banking collapse and the Depression of the 1930s. There were also accusations
of fraud or conflict of interest in commercial bank marketing of the securities they
underwrote. In addition to the separation of commercial and investment banking,
Glass argued for Regulation Q (which prohibited the payment of interest on de-
mand deposits) as a way to discourage banks from depositing funds in other banks.
Ironically, recent empirical evidence points to stabilizing effects from combining
commercial and investment banking prior to the Depression, and there is no evid-
ence of conflict of interest in doing so (Kroszner and Rajan, 1994). White (1986)
found that banks with securities affiliates had lower probabilities of failure and
enjoyed greater diversification of earnings. Kroszner and Rajan (1994) found that
the ex post performance of investments in securities underwritten by banks were
at least as good as for other securities. Benston (1989) argues that there was never
really any evidence unearthed by Congress that pointed to a destabilizing in-
fluence from underwriting or to a pattern of conflict of interest. Congress never
bothered with the evidence. The coincidence of the banking collapse of 1931-
1933 and the stock market collapse of 1929 provided all the evidence Congress
needed to make its decision.
While Glass did not support the creation of federal deposit insurance, his Sen-
ate Banking Committee's Pecora hearings served to buttress the case for deposit
insurance made by his opponents, Congressman Henry Steagall and Senator Huey
Long. They used the Pecora hearings as evidence that a banking system based on
large-scale banks was prone to crisis, arguing that commercial banks in financial
centers were the source of excessive speculation that had led to the Depression.
Unit bankers, in contrast, were-like the public-portrayed as the victims of
speculative excess. Steagall, Long, and their allies succeeded in making banking
398 MACROECONOMETRICS
reform a central issue in the forum of public debate. Despite the enormous losses
and evidence of moral-hazard problems in state-level deposit insurance systems of
the 1920s (White, 1983; Calomiris, 1990, 1992, 1993b), and despite the opposition
of Glass, Roosevelt, the Federal Reserve, and the Treasury, advocates of deposit
insurance, and capital assistance to banks through the Reconstruction Finance
Corporation, pushed though legislation resuscitating unit banks and effectively
restricting bank consolidation (Flood, 1991; Calomiris and White, 1994). Steagall's
success was one of the most impressive examples of legislative maneuvering in
the history of congressional politics.
To sum up, economic doctrine and regulation are important state variables for
determining macroeconomic outcomes. The Great Depression illustrates how false
beliefs and wrong-headed policies can emerge during times of crisis. Crisis-
induced changes often are not clearly thought through and may reflect motives that
have more to do with political opportunism than with social welfare. III-advised
institutional changes can have large and persistent costs because of the difficulty
of reversing them. The persistence of inefficient institutions can reflect the opera-
tion of special interests that defend the institution (as in unit banking) or the
protracted process of social learning about the costs of institutions (as in Fed
targeting in the interwar period, and the rigid adherence to gold standard rules in
many countries during 1929-1933).
Econometric Lessons
suspension, and the current fiduciary standard). Less clearcut, more gradual in-
stitutional changes have been important as well-for example, Fed learning about
monetary targeting during the Great Depression, and the movement away from a
unit banking system in the 1980s and 1990s.
The institutional history of a nation is difficult to define as an aggregate vari-
able, much less to control for when analyzing time series aggregates. Such a task
is particularly difficult given the frequent changes in institutions that take place.
But in disaggregated data, and in comparisons across countries, one can exploit
cross-sectional variation to test for, and control for, the importance of institutional
effects. In this respect, the methodological lessons of institutional non-neutrality
are quite similar to those of long swings.
Conclusion
In what ways do the facts of the past confirm and challenge macroeconomists'
existing models of the economy today? Available statistical evidence on the
cyclical properties of prices, wages, and other variables, as well as historical evid-
ence on the pace of technological change, are consistent with traditional explana-
tions of business cycles that revolve around shocks to aggregate demand as the
most important sources of economic fluctuation.
That does not mean that mainstream macroeconomic models based on sticky
wages and prices, or statistical analyses of time-series aggregates suggested by
those models, provide a complete explanation of American macroeconomic his-
tory. Perhaps surprisingly, the new frontiers of macroeconomics remain the old
frontiers that have been beckoning for half a century. In particular, economists
should devote more effort to understanding the history of low-frequency cycles
and severe, long-lived depressions.
Despite all that we have learned about the history of the economy and about the
sources of shocks during the 1930s in particular, macroeconomists still lack a fully
satisfying explanation of the protracted duration of high unemployment and excess
capacity during the Depression. More generally, standard macroeconomic analy-
sis, which focuses on high-frequency cycles produced by transitory demand shocks,
has neglected the phenomena surrounding Kuznets cycles-cycles of eighteen-
year duration that are associated with waves of new settlement, construction
activity, city building, and migration.
The lack of attention to low-frequency cycles is an important gap in our know-
ledge about the economy. Anyone attempting to provide a concise summary of
American economic development will find himself essentially plotting out the
Kuznets cycle. During the nineteenth and early twentieth centuries, especially, as
America pushed back the physical frontier westward, building cycles were central
400 MACROECONOMETRICS
(1993), Wallis (1989), Bresnahan and Raff (1991), Bernanke and James (1991),
Calomiris and Hubbard (1994), and Bernanke (1994) on the Great Depression
exemplify such a research agenda. All of these papers have in common the use of
cross-sectional variation to identify important macroeconomic shocks and pro-
pagators during the Depression. Margo and Wallis focus on the question of how
employment supply shifted in response to New Deal programs. Bresnahan and
Raff show how the Depression produced important endogenous technological
responses to the collapse of aggregate demand. Bernstein argues that cross-
sectional variation in the performance of different sectors is consistent with
significant changes in consumer demand brought on by the Depression. Bernanke
and James and Calomiris and Hubbard use cross-country and cross-firm differences
to link debt deflation and capital market constaints to the collapse of investment.
Macroeconomic history also teaches us that economic institutions are an inte-
gral part of macroeconomic history and important contributors to path-dependent
economic change. The Federal Reserve System is a particularly important
example. Like other important changes in institutions, its origins can be traced to
an adverse macroeconomic event. The Fed fell short of achieving the objectives
it was designed to achieve because its powers and the structure of the banking
system were constrained by other institutional arrangements-the gold standard
and preexisting unit banking. Over time, and in response to other events, the
structure or operating rules of the Fed have changed dramatically (for example, in
1935, 1951, and 1979), often with significant macroeconomic consequences.
Trying to model and measure the impact of institutional change, and test ex-
planations of institutional change, are challenges that have mainly been ignored by
macroeconomists and macroeconometricians. "Endogenizing" institutions and
tracing the long-run macroeconomic implications of specific institutional changes
pose formidable challenges to macroeconomic theory and measurement; but the
alternative-pretending that institutional change is an irrelevant sideshow-has
nothing to recommend it except the bliss of ignorance.
Donald McCloskey sometimes argues that Ph.D. students are the only audience
of economists worth addressing. With that advice in mind, we conclude with a list
of recommendations for students willing to brave the frontiers of macroeconomic
history. (1) Interesting questions about macroeconomic history often flow from an
understanding of the lives of real people. It does not hurt to become acquainted
with facts other than statistics. (2) Research that addresses questions posed by
models is not usually as interesting as research that addresses questions posed by
historical data or events. (3) Searching history for controlled experiments to test
models is a fine occupation, so long as the history surrounding the experiments is
not viewed as incidental to the exercise. (4) Scholars should not be afraid to tell
and defend stories on grounds of plausibility. Stories are no substitute for logic or
facts, but without them, economics is a rudderless ship. Researchers should look
402 MACROECONOMETRICS
Acknowledgments
The authors thank Barry Eichengreen, Kevin Hoover, Robert Margo, Allan Meltzer,
Kenneth Snowden, Peter Temin, and Jeffrey Williamson for helpful discussions.
Notes
as the railroads were able to cooperate, it was profitable for development to follow a regular, deliberate
process. Once cooperation broke down, the only way to seize market share was to build ahead of
demand.
11. Even more fundamentally, scholars disagree over whether long swings reflected an endog-
enous cyclical process-in which cycles of similar duration repeated naturally-or a shock-and-
propagation process, in which the timing and duration of the cycles reflected unique and unpredictable
disturbances. Easterlin's (1966) emphasis on immigration led him to argue that "echo effects" may
have explained the average duration of the cycle-a wave of immigrants would produce a new gen-
eration of frontier builders roughly at eighteen-year intervals.
12. That is not to say that investment banking was unimportant in the United States as a means
of financing industry. But despite the efforts and successes of investment banks in reducing some
firms' costs of external finance (DeLong, 1991), such financing was confined to the largest, most
mature firms in the economy, and new issues typically were restricted to senior claims against those
firms (bonds and preferred stock) because of the high cost of placing common stock (Calomiris and
Raff, 1994).
13. To some extent, especially in New York, deposits on foreign banks, wired via the trans-
Atlantic cable, may have been 'a cheaper means to introduce reserves to the U.S. banking system. But
as Calomiris and Hubbard (l994b) point out, there were limitations to the substitutability between
foreign bank balances and gold, particularly given the risk of runs on banks.
14. This was not true of the waves of banking failures and bank runs during the Great Depression
(Wicker. 1993). These began with regional problems in the South and Midwest, associated with severe
deflation and exposure by those regions' banks to loan losses. Only later did problems spread to
financial centers in the East and abroad. Great Depression banking problems also occurred near a
business-cycle trough, while national banking era panics coincided with cyclical peaks.
15. It is important to note that many, possibly most, bank failures during the 1930s did not coincide
with runs or panics. Many banks failed as the result of isolated insolvencies in the wake of continuing
deterioration in their panicular assets. Panics or runs did occur, nationally and regionally, during the
19305, but only as occasional punctuation points in the general trend of asset value decline and failure.
Thus it would be premature, and probably incorrect, to attribute most of the bank failures of the 19305
to panic episodes. For recent research trying to measure the importance of general runs on banks for
producing bank failure, see Wicker (1993) and Calomiris and Mason (1994).
16. In the face of new competitive pressures on American banks, and the failures of many banks
and thrifts in the 19805, branching limitations and limits on the powers of commercial banks have been
relaxed recently. In many ways, this process mirrors the branching and merger wave in banking during
the 19205. This illustrates how regulatory reform typically requires extreme circumstances.
References
Abramovitz, Moses. (1964). Evidence ofLong Swings in Aggregate Construction since the
Civil War. New York: Columbia University Press.
Allen, Steven G. (1992). "Changes in the Cyclical Sensitivity of Wages in the United
States, 1891-1987." American Economic Review 82, March, 122-140.
Alogoskoufis, George S., and Smith, Ron. (1991). "The Phillips Curve, the Persistence of
Inflation, and the Lucas Critique: Evidence from Exchange-Rate Regimes." American
Economic Review 81, December, 1254-1275.
Atack, Jeremy, and Fred Bateman. (1994). "Did the United States Industrialize Too Slowly?"
Working Paper, Vanderbilt University, July.
404 MACROECONOMETRICS
Atack, Jeremy, Fred Bateman, and Thomas Weiss. (1980). "The Regional Diffusion and
Adoption of the Steam Engine in American Manufacturing." Journal of Economic
History 40, June, 281-308.
Azariadis, Costas. (1975). "Implicit Contracts and Underemployment Equilibria." Journal
of Political Economy 83, December, 1183-1202.
Azariadis, Costas. (1983). "Implicit Contracts and Fixed Price Equilibria." Quarterly Jour-
nal of Economics 98, Supplement, 1-22.
Backus, David K. and Patrick J. Kehoe. (1992). "International Evidence on the Historical
Properties of Business Cycles." American Economic Review 82, September, 864-888.
Balke, Nathan S. and Robert J. Gordon. (1989). "The Estimation of Prewar Gross National
Product: Methodology and New Evidence." Journal ofPolitical Economy 97, February,
38-92.
Barro, Robert 1. (1979). "Second Thoughts on Keynesian Economics." American Economic
Review 69, May, 54-59.
Barro, Robert J. (1981). "Output Effects of Government Purchases." Journal of Political
Economy 89, December, 1086-1121.
Barro, Robert J. (1987). "Government Spending, Interest Rates, Prices, and Budget Deficits
in the United Kingdom, 1701-1918." Journal of Monetary Economics 20, September,
221-248.
Barro, R.I., and R.G. King. (1984). "Time-Separable Preferences and Intertemporal Sub-
stitution Models of the Business Cycle." Quanerly Journal of Economics 99, Novem-
. ber, 817-839.
Barsky, Robert B., N. Gregory Mankiw, Jeffrey A. Miron, and David N. Weill. (1988).
'The Worldwide Change in the Behavior of Interest Rates and Prices in 1914." Euro-
pean Economic Review 32, 1123-1154.
Benhabib, Jess, Richard Rogerson, and Randall Wright. (1991). "Homework in Macro-
economics: Household Production and Aggregate Auctuations." Journal of Political
Economy 99, 1166-1187.
Benston, George J. (1989). The Separation of Commercial and Investment Banking: The
Glass-Steagall Act Revisited and Reconsidered. Norwell, MA: Kluwer.
Berle, Adolph A., and Gardiner Means. (1932). The Modem Corporation and Private
Property. New York: Columbia University Press.
Bemanke, Ben S. (1983). "Nonmonetary Effects of the Financial Crisis in the Propagation
of the Great Depression." American Economic Review 73, June, 257-276.
Bernanke, Ben S. (1994). "The Macroeconomics of the Great Depression: A Comparative
Approach." NBER Working Paper No. 4814, August.
Bernanke, Ben S., and Harold James. (1991). "The Gold Standard, Deflation, and Financial
Crisis in the Great Depression." In R. Glenn Hubbard (ed.), Financial Markets and
Financial Crises (pp. 33-68). Chicago: University of Chicago Press.
Bemanke, Ben S., and Martin L. Parkinson. (1991). "Procyclical Labor Productivity and
Competing Theories of the Business Cycle: Some Evidence from Interwar U.S. Manu-
facturing Industries." Journal of Political Economy 99, 439-459.
Bernanke, Ben S. and James L. Powell. (1986). "The Cyclical Behavior of Industrial Labor
Markets: A Comparison of the Prewar and Postwar Eras." In Robert 1. Gordon (ed.), The
HISTORICAL MACROECONOMETRICS 405
American Business Cycles: Continuity and Change (pp. 583-638). Chicago: University
of Chicago Press.
Bernstein, Michael A. (1987). The Great Depression: Delilyed Recovery and Economic
Change in America 1929-1939. Cambridge: Cambridge University Press.
Bils, Mark. (1989). "Pricing in a Customer Market." Quarterly Journal of Economics 104,
November, 699-718.
Blanchard, Olivier, and Nobuhiro Kiyotaki. (1987). "Monopolistic Competition and the
Effects of Aggregate Demand." American Economic Review 77, September, 647-666.
Blanchard, Olivier 1., and Danny Quah. (1988). 'The Dynamic Effects of Aggregate Demand
and Supply Disturbances." Working Paper, Massachusetts Institute of Technology.
Blanchard, Olivier J., and Lawrence H. Summers. (1988). "Why Is Unemployment So High
in Europe? Beyond the Natural Rate Hypothesis." American Economic Association
Papers and Proceedings, May, 182-187.
Bodenhorn, Howard. (1992). "Capital Mobility and Financial Integration in Antebellum
America." Journal of Economic History 52, September, 585-610.
Bordo, Michael D., and Barry Eichengreen (eds.). (1993). A Retrospective on the Bretton
Woods System: Lessons for International Monetary Refonns. Chicago: University of
Chicago Press.
Bordo, Michael D., and Finn Kydland. (1990). "The Gold Standard as a Rule." Working
Paper, Rutgers University.
Bowley, A.L. (1904). "Tests of National Progress." Economic Journal 14, September, 457-
465.
Bresnahan, Timothy F., and Daniel M.G. Raff. (1991). "Intra-industry Heterogeneity and
the Great Depression: The American Motor Vehicles Industry, 1929-1935." Journal of
Economic History 51, June, 317-331.
Bullard, James B. (1994). "How Reliable Are Inflation Reports?" Monetary Trends, Federal
Reserve Bank of St. Louis, February.
Bums, Arthur F. (1951). "Mitchell on What Happens During Business Cycles." In Con-
ference on Business Cycles (pp. 3-33). New York: National Bureau of Economic
Research.
Cagan, PhiIlip. (1975). "Changes in the Recession Behavior of Wholesale Prices in the
I920s and Post-World War II." Explorations in Economic Research 2, Winter, 54-104.
Calomiris, Charles W. (1988a). "Institutional Failure, Monetary Scarcity, and the Depre-
ciation of the Continental." Journal of Economic History 48, March, 47-68.
Calomiris, Charles W. (1988b). "Price and Exchange Rate Determination During the
Greenback Suspension." Oxford Economic Papers 40, December, 719-750.
Calomiris, Charles W. (1990). "Is Deposit Insurance Necessary? A Historical Perspective."
Journal of Economic History 50, June, 283-295.
Calomiris, Charles W. (1992). "Do 'Vulnerable' Economies Need Deposit Insurance:
Lessons from the U.S. Agriculture in the I920s." In Philip L. Brock (ed.), lfTexas Were
Chile: A Primer on Banking Refonn (pp. 237-315,450-458). San Francisco: Institute
for Contemporary Studies.
Calomiris, Charles W. (1993a). "Greenback Resumption and Silver Risk: The Economics
and Politics of Monetary Regimes Change in the United States, 1862-1900." In M.D.
406 MACROECONOMETIUCS
Bordo and Forrest Capie (eds.), Monetary Regimes in Transition (pp. 86-132). Cam-
bridge: Cambridge University Press.
Calomiris, Charles W. (1993b). "Regulation, Industrial Structure, and Instability in U.S.
Banking: An Historical Perspective." In Michael Klausner and Lawrence J. White (eds.),
Structural Change in Banking (pp. 19-116). Homewood, IL: Business-One Irwin.
Calomiris, Charles W. (1993c). "Financial Factors in the Great Depression." Journal of
Economic Perspectives 7, Spring, 61-85.
Calomiris, Charles W. (1994). "The Costs of Rejecting Universal Banking: American
Finance in the German Mirror, 1870-1914." In Naomi Lamoreaux and Daniel Raff
(eds.), Coordination and Information: Historical Perspectives on the Organization of
Enterprise. Chicago: University of Chicago Press.
Calomiris, Charles W., and Gary Gorton. (1991). "The Origins of Banking Panics: Models,
Facts, and Bank Regulation." In R. Glenn Hubbard (ed.), Financial Markets and Finan-
cial Crises (pp. 109-173). Chicago: University of Chicago Press.
Calomiris, Charles W., and Christopher Hanes. (1994). "Consistent Output Series for the
Antebellum and Postbellum Periods: Issues and Preliminary Results." Journal of Eco-
nomic History 54, June, 409-422.
Calomiris, Charles W., and R. Glenn Hubbard. (1989). "Price Aexibility, Credit Availability,
and Economic Auctuations: Evidence from the United States, 1894-1909." Quarterly
Journal of Economics 104, February, 429-452.
Calomiris, Charles W., and R. Glenn Hubbard. (I 994a). "Internal Finance and Invest-
ment: Evidence from the Undistributed Profits Tax of 1936-1937." Journal ofBusiness,
forthcoming.
Calomiris, Charles W., and R. Glenn Hubbard. (1994b). "International Adjustment under
the Classical Gold Standard: Evidence from the U.S. and Britain, 1879-1914." In
Bayoumi and Eichengreen (eds.), Modern Perspectives on the Gold Standard, Cam-
bridge: Cambridge University Press, forthcoming.
Calomiris, Charles W., and Joseph R. Mason. (1994). "Contagion and Bank Failures during
the Depression: The June 1932 Chicago Banking Panic." NBER/No. 4934, November.
Working Paper.
Calomiris, Charles W., and Raff, Daniel M.G. (1994). "The Evolution of Market Structure,
Information, and Spreads in American Investment Banking." In M.D. Bordo and
R. Sylla (eds.), Anglo-American Finance. Homewood, IL: Irwin.
Calomiris, Charles W., and Larry Schweikart. (1991). "The Panic of 1857: Origins, Trans-
mission, and Containment." Journal of Economic History 51, December, 807-834.
Calomiris, Charles W., and Eugene N. White. (1994). "The Origins of Federal Deposit
Insurance." In Claudia Goldin and Gary D. Libecap (eds.), The Regulated Economy:
A Historical Approach to Political Economy (pp. 145-188). Chicago: University of
Chicago Press.
Calvo, Guillermo A. and Edmund S. Phelps. (1977). "Employment-Contingent Wage
Contracts." Journal of Monetary Economics, Supplement, 160-168.
Carlton, Dennis W. (1986). "The Rigidity of Prices." American Economic Review 76,
September, 637-658.
Carosso, Vincent, P. (1970). Investment Banking in American: A History. Cambridge, MA:
Harvard University Press.
HISTORICAL MACROECONOMETRICS 407
Carter, Susan B. and Richard Sutch. (1990). "The Labour Market in the 1890's: Evidence
from Connecticut Manufacturing." In Erik Aerts and Barry Eichengreen (eds.), Un-
employment and Underemployment in Historical Perspective. Studies in Social and
Economic History, Vol. 12. City: Leuven University Press.
Chandler, Alfred D. (1972). "Anthracite Coal and the Beginnings of the Industrial Revo-
lution in the United States." Business History Review 46, Summer, 141-181.
Chandler, Alfred D. (1977). The Visible Hand: The Managerial Revolution in American
Business. Cambridge, MA: Harvard University Press.
Christiano, Lawrence J., and Martin Eichenbaum. (1992). "Current Real-Business-Cycle
Theories and Aggregate Labor-Market Auctuations." American Economic Review 82,
June, 430-450.
Clark, Truman A. (1986). "Interest Rate Seasonals and the Federal Reserve." Journal of
Political Economy 94, February, 76-125.
Cooley, Thomas F., and Lee E. Ohanian. (1991). "The Cyclical Behavior of Prices."
Journal of Monetary Economics 28, August, 25-60.
Crucini, Mario J. (1994). "Sources of Variation in Real Tariff Rates: The United States,
1900-1940." American Economic Review 84, June, 732-744.
David, Paul A. (1975). Technical Choice, Innovation, and Economic Growth: Essays on
American and British Experience in the Nineteenth Century. New York: Cambridge
University Press.
David, Paul A. (1990). "The Dynamo and the Computer: An Historical Perspective on the
Modern Productivity Paradox." American Economic Review 80, May, 355-361.
David, Paul A. (1991). "Economic History Association Presidential Address." Working
Paper, Stanford University.
David, Paul A., and Joshua L. Rosenbloom. (1990). "Marshallian Factor Market Exter-
nalities and the Dynamics of Industrial Location." Journal of Urban Economics 28,
349-370.
Davis, Lance E. (1957). "Sources of Industrial Finance: The American Textile Industry, A
Case Study." Explorations in Entrepreneurial History 9, 190-203.
Davis, Lance E. (1960). "The New England Textile Mills and the Capital Markets: A Study
of Industrial Borrowing." Journal of Economic History 20, 1-30.
Davis, Lance, E. (1963). "Capital Immobilities and Finance Capitalism: A Study of
Economic Evolution in the United States." Explorations in Entrepreneurial History 15,
Fall, 88-105.
Davis, Lance E. (1965). "The Investment Market, 1870-1914: The Evolution of a National
Market." Journal of Economic History 25, September, 355-393.
Dawley, Allen. (1976). Class and Community: The Industrial Revolution in Lynn Massa-
chusetts. Cambridge, MA: Harvard University Press.
DeLong, J. Bradford. (1991). "Did J.P. Morgan's Men Add Value? An Economist's Per-
spective on Financial Capitalism." In Peter Temin (ed.), Inside the Business Enterprise:
Historical Perspectives on the Use of Information (pp. 205-236). Chicago: University
of Chicago Press.
DeLong, J. Bradford, and Lawrence Summers. (1986) "The Changing Cyclical Variability of
Economic Activity in the United States." In Robert J. Gordon (ed.), The American Busi-
ness Cycle: Continuity and Change (pp. 679-732). Chicago: University of Chicago Press.
408 MACROECONOMETRICS
Friedman, Milton. (1952). "Price, Income and Monetary Changes in Three Wartime Periods."
American Economic Review, May, - .
Friedman, Milton, and Anna Jacobson Schwartz. (1963). A Monetary History ofthe United
States 1867-1960. Princeton, NJ: Princeton University Press.
Gallman, Robert. (1960). "Commodity Output, 1839-1899." In Trends in the American
Economy in the Nineteenth Century, NBER Studies in Income and Wealth (Vol. 24) (pp.
13-71). New York: Columbia University Press.
Gallman, Robert. (1966). "Gross National Product in the United States, 1834-1909." In
Output, Employment and Productivity in the United States After 1800, NBER Studies
in Income and Wealth (Vol. 30) (pp. 3-90). New York: Columbia University Press.
Goldin, Claudia, and Stanley Engerman. (1993). "Seasonality in Nineteenth Century
American Labor Markets." In Donald Schaefer and Thomas Weiss (eds.), Economic
Development in Historical Perspective. Stanford, CA: Stanford University Press.
Goldin, Claudia, and Robert Margo. (1992). "Wages, Prices, and Labor Markets Before the
Civil War." In Claudia Goldin and Hugh Rockoff (eds.), Strategic Factors in Nineteenth
Century American Economic Development. Chicago: University of Chicago Press.
Goldin, Claudia, and Kenneth Sokoloff. (1984). "The Relative Productivity Hypothesis of
Industrialization: The American Case, 1820 to 1850." Quarterly Journal of Economics
69, August, 461-487.
Gordon, Robert J. (1981). "Output Fluctuations and Gradual Price Adjustment." Journal of
Economic Literature 19, June, 493-530.
Gordon, Robert J. (1986). "Introduction: Continuity and Change in Theory, Behavior, and
Methodology." In Robert J. Gordon (ed.), The American Business Cycle: Continuity and
Change (pp. 1-33). Chicago: University of Chicago Press.
Gordon, Robert 1. (1990). "What Is New-Keynesian Economics?" Journal of Economic
Literature 28, September, 1115-1171.
Greenwood, Jeremy Z. Hercowitz, and G.W. Huffman. (1988). "Investment, Capacity
Utilization. and the Real Business Cycle:' American Economic Revie"' 78. June. 402-
417.
Grossman. Richard S. (1993). "The Macroeconomic Consequences of Bank Failures Under
the National Banking System:' Explorations in Economic History 30. July. 294-320.
Hanes. Christopher. (1993). 'The Development of Nominal Wage Rigidity in the Late
Nineteenth Century:' American Economic Re~'iew 83. September. 732-756.
Hanes. Christopher. (1994). "Changes in the Cyclical Behavior of Nominal Prices. 1869-
1990." Working Paper. University of Pennsylvania.
Hansen. Gary D.. and Edward C. Prescott. (1993). "Did Technology Shocks Cause the
1990-1991 Recession?" American Economic Review May. 280-286.
Harley. C. Knick. (1982). "Oligopoly Strategy and the Timing of American Railroad
Construction." Journal of Economic HisTory 42. December. 797-824.
Hatton. Timothy 1.. and Jeffrey G. Williamson. (1991). "Wage Gaps Between Farm and
City: Michigan in the I890s:' ExploraTions in Economic History 28. October. 381-408.
Hekman. John S. (1980). "The Product Cycle and New England Textiles:' Quarterly Jour-
nal of Economics 95. June. 697-717.
Helms. Lloyd A. (1939). The Contributions of Lord OversTone 10 the Theory of Currency
and Banking. Urbana: University of Illinois Press.
410 MACROECONOMETRICS
Hicks, John R. (1937). "Mr. Keynes and the 'Classics': A Suggested Interpretation."
Econometrica 5, 147-159.
Hughes, Jonathan. (1990). American Economic History (3rd ed.). Glenview, IL: Scott,
Foresman.
Hughes, Jonathan. (1991). The Governmental Habit Redux: Economic Controls from
Colonial Times to the Present. Princeton, NJ: Princeton University Press.
Hultgren, Thor. (1960). "Changes in Labor Cost During Cycles in Production and Busi-
ness." National Bureau of Economic Research Occasional Paper 74.
Jacoby, Sanford. (1985). Employing Bureaucracy: Managers. Unions, and the Transforma-
tion of Work in American Industry. 1900-1945. New York: Columbia University Press.
Jacoby, Sanford and Sunil Sharma. (1992). "Employment Duration and Industrial Labor
Mobility in the United States. 1880-1980." Journal of Economic History 52, March,
161-179.
James, John A. (1978). Money and Capital Markets in Postbellum America. Princeton, NJ:
Princeton University Press.
James. John A. (1989). "The Stability of the Nineteenth-Century Phillips Curve Relation-
ship." Explorations in Economic History 26, April, 117-134.
James, John A. (1993). "Changes in Economic Instability in Nineteenth-Century America."
American Economic Review 83, September, 710-731.
Katz, Lawrence F., and Lawrence H. Summers. (1989). "Industry Rents: Evidence and
Implications." Brookings Papers: Microeconomics, 209-290.
Keynes, John Maynard. (1931/1953). A Treatise on Money. Vol. 2, The Applied Theory of
Money. London: Macmillan.
Keynes, John Maynard. (1936/1964). The General Theory of Employment. Interest. and
Money. New York: Harcourt Brace Jovanovich.
Keynes, John Maynard. (1939). "Relative Movements of Real Wages and Output." Eco-
nomic Journal 49, March, 35-51.
Keyssar, Alexander. (1986). Out of Work: The First Century of Unemployment in Massa-
chusetts. New York: Cambridge University Press.
Khan, B. Zorina, and Kenneth L. Sokoloff. (1993). '''Schemes of Practical Utility': Entre-
preneurship and Innovation Among 'Great Inventors' in the United States, 1790-1865."
Journal of Economic History 53, June, 289-307.
Kindleberger, Charles P. (1973). The World in Depression. 1929-1939. Barkeley: Univer-
sity of California Press.
Kroszner, Randall S., and Raghuram Rajan. (1994). "Is the Glass-Steagall Act Justified? A
Study of the U.S. Experience with Universal Banking Before 1933." American Eco-
nomic Review 84, September, 810-832.
Krueger, Alan B., and Lawrence H. Summers. (1987). "Reflections on the Inter-industry
Wage Structure." In Kevin Lang and Jonathan S. Leonard (eds.), Unemployment and the
Structure of Labor Markets (pp. 17-47). New York: Basil Blackwell.
Krugman, Paul. (1991). "Increasing Returns and Economic Geography." Journal of Polit-
ical Economy 99, June, 483-499.
Kuznets, Simon. (1951). "Comment on Joseph Schumpeter." In Conference on Business
Cycles (pp. 155-162). New York: National Bureau of Economic Research.
HISTORICAL MACROECONOMETRICS 411
Kydland, Finn, and Edward Prescott. (1990). "Business Cycles: Real Facts and a Monetary
Myth." Federal Reserve Bank of Minneapolis Quarterly Review 14, Spring, 3-18.
Lamoreaux, Naomi R. (1985). The Great Merger Movement in American Business, 1895-
1904. Cambridge: Cambridge University Press.
Lamoreaux, Naomi R. (1994). Insider Lending: Banks, Personal Connections, and Eco-
nomic Development in Industrial New England. Cambridge: Cambridge University Press.
Lazear, Edward. (1981). "Agency, Earnings Profiles, Productivity and Hours Restrictions."
American Economic Review 71, June, 606-620.
Lebergott, Stanley. (1964). Manpower in Economic Growth: The American Record Since
1800. New York: McGraw-Hili.
Lerdau, E. (1957). "On the Measurement of Tariffs: The U.S. Over Forty Years." Economia
1nternazionale 10, May, 232-244.
Lindbeck, Assar, and Dennis J. Snower. (1988). The Insider-Outsider Theory of Employ-
ment and Unemployment. Cambridge, MA: MIT Press.
Lipsey, Robert E. (1963). Price and Quantity Trends in the Foreign Trade of the United
States. Princeton, NJ: Princeton University Press.
Long, Clarence D. (1960). Wages and Earnings in the United States, 1860-1890. Princeton,
NJ: Princeton University Press.
Lucas, Robert E. (1972). "Expectations and the Neutrality of Money." Journal ofEconomic
Theory 4, April, 103-124.
Lucas, Robert E. (1976). "Econometric Policy Evaluation: A Critique." Carnegie-Rochester
Series on Public Policy 1.
Lucas, Robert E., and Michael Woodford. (1993). "Real Effects of Monetary Shocks in an
Economy with Sequential Purchases." NBER Working Paper No. 4250, January.
Mankiw, N. Gregory. (1985). "Small Menu Costs and Large Business Cycles: A
Macroeconomic Model of Monopoly." QuarteriyJournal ofEconomics 100, May, 529-
539.
Mankiw, N. Gregory. (1989). "Real Business Cycles: A New Keynesian Perspective."
Journal of Economic Perspectives 3, Summer, 79-90.
Mankiw, N. Gregory. (1990). "A Quick Refresher Course in Macroeconomics." Journal of
Economic Literature 28, December, 1645-1660.
Mansfield, Edwin. (1961). "Technical Change and the Rate oflmitation." Econometrica 29,
October, 741-766.
Mansfield, Edwin. (1983). "Long Waves and Technological Innovation." American Eco-
nomic Association Papers and Proceedings 73, May, 141-145.
Manuelli, Rodolfo E. (1986). "Modem Business Cycle Analysis: A Guide to the Prescott-
Summers Debate." Federal Reserve Bank ofMinneapolis Quarterly Review 10, Fall, 3-
8.
Margo, Robert A. (1990a). "Unemployment in 1910: Some Preliminary Findings." In Erik
Aerts and Barry Eichengreen (eds.), Unemployment and Underemployment in llis-
torical Perspective, Studies in Social and Economic History Vol. 12. City: Leuven
University Press.
Margo, Robert A. (1990b). "The Incidence and Duration of Unemployment: Some Long-
Term Comparisons." Economics Letters 32, January, 217-220.
412 MACROECONOMETRICS
Officer, Lawrence H. (1986). "The efficiency of the Dollar-Sterling Gold Standard." Jour-
nal of Policial Economy 94, October, 1038-1073.
Olmstead, Alan, and Paul, Rhode. (1985). "Rationing Without Government The West
Coast Gas Famine of 1920." American Economic Review 75, December, 1044-1055.
Ozanne, Robert. (1968). Wages in Practice and Theory: McCormick and International
Harvester, 1860-1960. Madison, WI: University of Wisconsin Press.
Plosser, Charles I. (1989). "Understanding Real Business Cycles." Journal of Economic
Perspectives 3, Summer, 51-77.
Raff, Daniel M.G. (1988). "Wage Determination Theory and the Five-Dollar Day at Ford."
Journal of Economic History 48, June, 387-399.
Ransom, Roger, Richard Sutch, and Susan Carter. (1992). "Reestimating the Annual Un-
employment Rate Series for the United States, 1890-1940." Research in Economic
History 14, 293-299.
Rees, Albert. (1961). Real Wages in Manufacturing, 1890-1914. Princeton, NJ: Princeton
University Press.
Romer, Christina. (l986a). "Spurious Volatility in Historical Unemployment Data." Jour-
nal of Political Economy 94, February, 1-37.
Romer, Christina. (1986b). "Is the Stabilization of the Postwar Economy a Figment of the
Data?" American Economic Review 76, June, 314-334.
Romer, Christina. (1989). "The prewar Business Cycle Reconsidered: New Estimatf$ of
Gross National Product, 1869-1908." Journal ofPolitical Economy 97, February, 1-37.
Romer, Christina. (1993). "The Nation in Depression." Journal of Economic Perspectives
7, Spring, 19-39.
Romer, Christina. (1994). "Remeasuring Business Cycles." Journal of Economic History
54, September, 573-609.
Romer, David. (1993). ''The New Keynesian Synthesis." Journal ofEconomic Perspectives
7, Winter, 5-22.
Rosenbloom, Joshua L. (1988). "Labor Market Institutions and the Geographic Integration
of Labor Markets in the Late Nineteenth Century United States." Ph.D. Dissertation,
Stanford University.
Rosenbloom, Joshua L. (1990). "One Market or Many? Labor Market Integration in the Late
Nineteenth-Century United States." Journal of Economic History 50, March, 85-107.
Rotemberg, Julio 1., and Michael Woodford. (1991). "Markups and the Business Cyde."
In Olivier Jean Blanchard and Stanley Fischer (eds.), NBER Macroeconomic Annual
1991. Cambridge, MA: MIT Press.
Sachs, Jeffrey. (1980). "The Changing Cyclical Behavior of Wages and Prices: 1890-
1976." American Economic Review 70, March, 78-90.
Salter, W.E.G. (1960). Productivity and Technical Change. Cambridge: Cambridge Uni-
versity Press.
Sargent, Thomas J. (198Ia). "The Ends of Four Big Inflations." Working Paper, University
of Minnesota.
Sargent, Thomas 1. (1981b). "Stopping Moderate Inflations: The Methods of Poincare and
Thatcher." Working Paper, University of Minnesota.
Schumpeter, Joseph A. (1939). Business Cycles: A Theoretical, Historical and Statisfical
Analysis of the Capitalistic Process. New York: McGraw-Hili.
414 MACROECONOMETRICS
Shapiro, Matthew D., and Mark W. Watson. (1988). "Sources of Business Cycle Fluctua-
tions." NBER Macroeconomics Annual 1988, 111-156.
Shaw, William Howard. (1947). Value of Commodity Output Since 1869. New York:
National Bureau of Economic Research.
Shergold, Peter R. (1975). "Wage Rates in Pittsburgh During the Depression of 1908."
Journal of American Studies 9, August 163-188.
Shister, Joseph. (1944). "A Note on Cyclical Wage Rigidity." American Economic Review
34, March, 111-116.
Simkovich, Boris A. (1994). "Agriculture and Industry in the New Republic: Explaining
America's Structural Transfonnation, 1800-1860." Working Paper, Harvard Univer-
sity, July.
Smith, Bruce D. (1985a). "Some Colonial Evidence on Two Theories of Money: Maryland
and the Carolinas." Journal of Political Economy 93, December, 1178-1211.
Smith, Bruce D. (1985b). "American Colonial Monetary Regimes: The Failure of the
Quantity Theory of Money and Some Evidence in Favor of an Alternate View." Cana-
dian Journal of Economics 18, August, 531-565.
Snowden, Kenneth. (1984). "Three Essays on the American Capital Market, 1870-1913."
Ph.D. Dissertation, University of Wisconsin-Madison.
Snowden, Kenneth. (1993). "The Evolution of Interregional Mortgage Lending Channels,
1870-1940." Working Paper, Harvard University.
Snowden, Kenneth, and Nidal Abu-Saba. (1994). "Why Did Late Nineteenth Century
American Mortgage Banking Fail?" Working Paper, Harvard University.
Sokoloff, Kenneth L. (1988). "Inventive Activity in Early Industrial America: Evidence
from Patent Records, 1790-1846." Journal of Economic History 48, December, 813-
850.
Solomou, Solornos. (1987). Phases of Economic Growth, 1850-1973: Kondratieff Waves
and Kuznets Swings. Cambridge: Cambridge University Press.
Solow, Robert M. (1968). "Distribution in the Long and Short Run." In Jean Marchal and
Bernard Ducros (eds.), The Distribution of National Income. New York: S1. Martin's
Press.
Spencer, Austin. (1977). "Relative Downward Industrial Price Flexibility, 1870-1921."
Explorations in Economic History 14, January, 1-19.
Stiglitz, Joseph D. (1986). "Theories of Wage Rigidities." In J. Butkiewicz (ed.), Keynes'
Economic Legacy: Contemporary Economic Theories. New York: Praeger.
Sundstrom, William A. (1990). "Was There a Golden Age of Flexible Wages? Evidence from
Ohio Manufacturing, 1892-1910." Journal of Economic History 50, June, 309-320.
Sundstrom, William A. (1992). "Rigid Wages or Small Equilibrium Adjustments? Evid-
ence from the Contraction of 1893." Explorations in Economic History 29, October,
430-455.
Sundstrom, William A., and Joshua L. Rosenbloom. (1993). "Occupational Differences in
the Dispersion of Wages and Working Hours: Labor Market Integration in the United
States, 1890-1903." Explorations in Economic History 30, October, 379-408.
Sylla, Richard. (1969). "Federal Policy, Banking Market Structure, and Capital Mobiliza-
tion in the United States, 1863-1913." Journal of Economic History 29, December,
657-686.
HISTORICAL MACROECONOMETRICS 415
Taussig, Frank William. (1931). A TariffHistory ofthe United States. New York: Putnam's.
Temin, Peter. (1989). Lessons from the Great Depression. Cambridge, MA: MIT Press.
Thomas, Brinley. (1954). Migration and Economic Growth. Cambridge: Cambridge Uni-
versity Press.
Tobin, James. (1993). "Price Flexibility and Output Stability: An Old Keynesian View."
Journal of Economic Perspectives 7, Winter, 45-65.
Towne, Marvin W., and Wayne D. Rasmussen. (1960). "Farm Gross Product and Gross
Investment in the Nineteenth Century." In Trends in the American Economy in the
Nineteenth Century (pp. 255-315). NBER Studies in Income and Wealth, Vol. 24. New
York: Columbia University Press.
United States Bureau of Labor Statistics. (1969). "Summary of Manufacturing Produ<:tion
Workers Earnings Series, 1939-1968." Bulletin 1616, Washington, DC.
United States Bureau of Labor Statistics. (1992). "Consumer Price Index Detailed Report."
March 1992.
United States Bureau of the Census. (1975). Historical Statistics of the United States.
Washington, DC: U.S. Government Printing Office.
United States Council of Economic Advisers. Various years. Economic Report ofthe Presi-
dent. Washington, DC: U.S. Government Printing Office.
von Tunzelmann, G.N. (1978). Steam Power and British Industrialization to 1860. Oxford:
Oxford University Press.
Wachter, Michael. (1970). "Cyclical Variation in the Interindustry Wage Structure." Ameri-
can Economic Review 60, March, 75-84.
Wallis, John 1. (1989). "Employment in the Great Depression: New Data and Hypotheses."
Explorations in Economic History 26, January, 45-72.
Warren, G.F., and F.A. Pearson. (1932). "Wholesale Prices in the United States for 135
Years, 1797 to 1932." Cornell University Agricultural Experiment Station, Memoir 142,
November.
Waston, Mark W. (1994). "Business-Cycle Durations and Postwar Stabilization of the U.S.
Economy." American Economic Review 84, March, 24-46.
Weinstein, Michael M. (1981). "Some Macroeconomic Impacts of the National Industrial
Recovery Act, 1933-1935." In Karl Brunner (ed.), The Great Depression Revisited.
Rochester Studies in Economics and Policy Issues, Vol. 2.
Weir, David R. (1986). ''The Reliability of Historical Macroeconomic Data for Comparing
Cyclical Stability." Journal of Economic History 46, June, 353-365.
Weir, David R. (1992). "A Century of U.S. Unemployment, 1890-1990: Revised Estimates
Evidence for Stabilization." Research in Economic History 14, 301-346.
Wheelock, David C. (1989). ''The Strategy, Effectiveness, and Consistency of Federal
Reserve Monetary Policy, 1924-1933." Explorations in Economic History 26, October,
453-476.
Wheelock, David C. (1990). "Member Bank Borrowing and the Fed's Contractionary
Monetary Policy During the Great Depression." Journal ofMoney, Credit and Banking
22, November, 409-426.
Wheelock, David c., and Subal C. Kumbhakar. (1994). "'The Slack Banker Dan<:es':
Deposit Insurance and Risk-Taking in the Banking Collapse of the 1920s." Explorations
in Economic History 31, July, 357-375.
416 MACROECONOMETRJCS
White, Eugene N. (1983). The Regulation and Reform of the American Banking System,
1900-1929. Princeton, NJ: Princeton University Press.
White, Eugene N. (1984). "Voting for Costly Regulation: Evidence from Banking Refer-
enda in Illinois, 1924." Southern Economic Journal 51, 1084-1098.
White, Eugene N. (1986). "Before the Glass-SteagaIl Act: An Analysis of the Investment
Banking Activities of National Banks." Explorations in Economic History 23, January,
33-55.
Wicker, Elmus. (1993). "A Reconstruction of the Gold and Banking Crises in 1931."
Working Paper, Indiana University, October.
Wood, Elmer. (1939). English Theories of Central Banking Control, 1819-1858. Cam-
bridge, MA: Harvard University Press.
Wright, Gavin. (1979). "Cheap Labor and Southern Textiles Before 1880." Journal of
Economic History 39, September, 655-680.
Wright, Gavin. (1981). "Cheap Labor and Southern Textiles, 1880-1930." Quarterly Jour-
nal of Economics 96, November, 605-629.
Wright, Gavin. (1986). Old South, New South. New York: Basic Books.
Wright, Gavin. (1990). "The Origins of American Industrial Success, 1879-1940." Ameri-
can Economic Review 80, September, 651-668.
Zarnowitz, Victor. (1992). Business Cycles: Theory, History, Indicators, and Forecasting.
Chicago: University of Chicago Press.
Commentary on Chapter 10
Steven N. Durlauf
Introduction
Charles Calomiris and Christopher Hanes have written a masterful survey of the
contribution of recent work in American economic history to our understanding
of macroeconomic theory. Underlying the various empirical analyses of the paper
is the argument that this work in macroeconomic history requires macroeconomists
to shift away from the standard conceptualization of macroeconomic data as the
realization of a time invariant stochastic process toward one that accounts for the
way in which the "cumulative past of the economy, including the history of shocks
and their effects, change the structure of the economy."
This comment is designed to complement Calomiris and Hanes by describing
what path dependence means from the perspective of statistics and economic
theory. The discussion is organized as follows. The following section provides a
couple of definitions of path dependence that represent ways of formally defining
the role of history in contemporaneous fluctuations. These definitions illustrate
how historical exercises of the type Calomiris and Hanes propose are an important
supplement to standard econometric exercises. The next two sections describe
recent work in macroeconomics which provides theoretical underpinnings for
Calomiris and Hanes's arguments on empirical methodology and extend the dis-
cussion of relevant economic theory to the notion of institutional evolution that
Calomiris and Hanes consider. The final section concludes.
Much of the motivation for Calomiris and Hanes's analysis stems from the belief
that history matters: "Some events have permanent importance for the future of
the economy." The idea that initial conditions, or some finite set of individual
events, can have long-run consequences for aggregate activity, which is often
called path dependence, may be given a rigorous probabilistic formulation. One
possible formal definition of path dependence is the following. Suppose that
aggregate output in the economy, Y" is an asymptotically stationary stochastic
process with a well-defined limit Y_. Without loss of generality, Y, can be taken
to be a function of the history of some underlying set of exogenous innovations
X,; denote the history of these innovations as F so that Y, = 'P(F,). Finally, let
"
417
418 MACROECONOMETRICS
This definition says that something in the history of the economy between 0 and
T has an effect on the long-run stochastic process characterizing aggregate activ-
ity. Equivalently, if aggregate output is path dependent according to this defini-
tion, then the process is nonergodic. Nonergodicity means that the limits of the
sample moments of the output process will depend on the particular sample path
realization of the economy. Since initial conditions can always be included in F r
- Fo, this definition represents a stochastic generalization of the sorts of multiple
equilibria which arise in deterministic models.
On the other hand, the ideas of positive feedbacks and self-reinforcing behavior
extend to a far larger set of models than those that exhibit path dependence as
defined by equation (I). For example, suppose that we take an economy that
exhibits multiple deterministic equilibria and superimpose on this model a selec-
tion mechanism that picks one of the equilibria each period based on the realiza-
tion of some random process. It is straightforward to see that all sample paths
generated by the model will have the same moments so long as the selection
mechanism is itself ergodic. Such an economy would not, by equation (I), be path
dependent, since the selection mechanism would now be part of the model. Dis-
tinguishing between the deterministic and stochastic versions of the model would
seem to be undesirable as both models possess the fundamental feature of self-
reinforcing behavior.
Such considerations suggest a second definition of path dependence that distin-
guishes models on the basis of whether shocks are or are not self-correcting.
Again, suppose that Y/ is a measurable function of some set of exogenous inno-
vations. Further, suppose that the stochastic process characterizing the exogenous
variables has the property that for all t > T, all innovations are equal to zero, which
means that no new shocks affect the economy after T. Aggregate output is path
dependent when the realized history of innovations affects the behavior of Y_ with
positive probability-that is,
Prob(Y_ I Fr) ::f:. Prob(Y_ I Fo) if X, = 0 if t > T. (2)
In other words, when all future innovations equal zero, then the realized sample
path of the economy can have long-run effects.
This definition of path dependence distinguishes those models in which the
history of shocks will eventually work their way out of an economy from those
models in which the history of shocks is not self-correcting. This definition incor-
porates stationary, nonergodic economies as well as economies that shift between
self-reinforcing regimes and therefore seems useful in characterizing the extent
COMMENTARY ON CHAPTER 10 419
One important feature of Calomiris and Hanes paper that has relevance for eco-
nomic theory concerns the endogeneity of institutions. Specifically, the authors
argue that a fundamental feature of economic development is the way in which
institutions such as the government learn to respond to differing economic
422 MACROECONOMETRICS
Conclusions
economic growth to business cycles to the emergence of ghettos. This recent work
in economic theory has, in tum, clarified the sort of path dependence that is likely
to emerge in data. Calorniris and Hanes provide a very valuable contribution to our
understanding of the empirical relevance of this theoretical work, as well as an im-
portant challenge to traditional methods of empirical work in macroeconomics by
demonstrating the importance of research in macroeconomic history in comple-
menting standard statistical analyses.
Acknowledgments
References
Arthur, W.B. (1989). "Increasing Returns, Competing Technologies and Lock-in by His-
torical Small Events: The Dynamics of Allocation Under Increasing Returns to Scale."
Economic Journal 99, 116-131.
Arthur, W.B. (1990). "Silicon Valley Locational Clusters: When Do Increasing Returns
Imply Monopoly?" Mathematical Social Sciences 19,235-251.
Azariadis, c., and A. Drazen. (1990). "Threshold Externalities and Economic Develop-
ment." Quarterly Journal of Economics 105, 501-526.
Benabou. R. (1993). "Workings of a City: Location, Education, and Production." Quanerly
Journal of Economics 108,619-652.
Benabou, R. (1994). "Education, Income Distribution, and Growth: The Local Connec-
tion." Working Paper No. 4798, National Bureau of Economic Research.
Brock, W. (1993). "Pathways to Randomness in the Economy: Emergent Nonlinearity in
Economics and Finance." Estudios Economicos 8, 3-55.
Bryant, J. (1985). "A Simple Rational Expectations Keynes-Type Model." QuanerlyJour-
nal of Economics 98, 525-529.
Cooper, R. (1987). "Dynamic Behavior of Imperfectly Competitive Economies with Mul-
tiple Equilibria." Working Paper No. 2388, National Bureau of Economic Research.
Cooper, R., and A. John. (1988). "Coordinating Coordination Failures in Keynesian Mod-
els." Quarterly Journal of Economics 103,441-465.
David, P. (1986). "Understanding the Economics of QWERTY: The Necessity of History."
In W. Parker (ed.), Economic History and the Modem Economist. Oxford: Basil
Blackwell.
David, P. (1988). "Path Dependence: Putting the Past in the Future of Economics." Work-
ing Paper, Stanford University.
Diamond, P. (1982). "Aggregate Demand in Search Equilibrium." Journal of Political
Economy 90, 881-894.
Durlauf, S.N. (1993). "Nonergodic Economic Growth." Review of Economic Studies 60,
349-366.
424 MACROECONOMETRICS
Durlauf, S.N. (l994a). "Path Dependence in Aggregate Output." Industrial and Corporate
Change 3, 173-198.
Durlauf, S.N. (1994b). "An Incomplete Markets Theory of Business Cycle fluctuations."
Working Paper, University of Wisconsin at Madison.
Durlauf, S.N. (l994c). "A Theory of Persistent Income Inequality." Working Paper, Uni-
versity of Wisconsin at Madison.
Durlauf, S.N. (l994d). "Neighborhood Feedbacks, Endogenous Stratification, and Income
Inequality." In W. Barnett, G. Gandolfo, and C. Hillinger (oos.), Dynamic Disequilibrium
Modelling: Proceedings ofthe Sixth International Symposium on Economic Theory and
Econometrics. Cambridge: Cambridge University Press.
Heller, W.P. (1986). "Coordination Failure Under Complete Markets with Applications
to Effective Demand." In W.P. Heller, R.M. Starr, and D.A. Starrett (eds.), Essays in
Honor of Kenneth J. Arrow (Vol. 2). Cambridge: Cambridge University Press.
Kremer, M., and E. Maskin. (1994). "Segregation of Workers by Skill and the Rise in
Inequality." Working Paper, Massachusetts Institute of Technology.
Krugman, P. (1991). "Increasing Returns and Economic Geography." Journal of Political
Economy 99, 483-499.
Krugman, P. (1993). "On the Number and Location of Cities." European Economic Review
37, 293-298.
Krugman, P. (1994). "fluctuations, Instability, and Agglomeration." Working Paper,
Stanford University.
Lucas, R.E. (1988). "On the Mechanics of Economic Development." Journal of Monetary
Economics 22,3-42.
Murphy, K., A. Shleifer, and R. Vishny. (1989). "Industrialization and the Big Push."
Journal of Political Economy 97, 1003-1026.
Romer, P. (1986). "Increasing Returns and Long Run Growth." Journal ofPolitical Economy
94, 1002-1037.
IIIFRONTIERS OF
MACROECONOMETRICS
11 MODELING VOLATILITY
DYNAMICS
Francis X. Diebold
Jose A. Lopez
Introduction
Good macroeconomic and financial theorists, like all good theorists, want to get
the facts straight before theorizing; hence, the explosive growth in the metho-
dology and application of time-series econometrics in the last twenty-five years.
Many factors fueled that growth, ranging from important developments in related
fields (see Box and Jenkins, 1970) to dissatisfaction with the "incredible identify-
ing restrictions" associated with traditional macroeconometric models (Sims, 1980)
and the associated recognition that many tasks of interest, such as forecasting,
simply do not require a structural model (see Granger and Newbold, 1979). A
short list of active subfields includes vector autoregressions, index and dynamic
factor models, causality, integration and persistence, cointegration, seasonality,
unobserved-components models, state-space representations and the Kalman filter,
regime-switching models, nonlinear dynamics, and optimal nonlinear filtering. Any
such list must also include models of volatility dynamics. Models of autoregressive
conditional heteroskedasticity (ARCH), in particular, provide parsimonious ap-
proximations to volatility dynamics and have found wide use in macroeconomics
and finance. 1 The family of ARCH models is the subject of this chapter.
Economists are typically introduced to heteroskedasticity in cross-sectional
427
428 MACROECONOMETRICS
B(L) = "2);L ,
i
~); < 00, bo = 1,
i=O ;=0
ARCH MODELS 429
1.8
,.
'6
02
005
004
x '0
2.5
'5
r....
1.8
000
004
·(1.00
TIM
"0
2.!
Figure 11.2. Daily logged swiss franc/dollar exchange rate (1974-1991) along
with the corresponding daily returns and squared daily returns
ARCH MODELS 431
E[e e) = {(J2 £
< 00, if t =r
r r 0, otherwise.
The uncorrelated innovation sequence {e,} need not be Gaussian and therefore
need not be independent. Nonindependent innovations are characteristic of
nonlinear time series in general and conditionally heteroskedastic time series in
particular.
In this section, we introduce the ARCH process within Wold's framework by
contrasting the polar extremes of the LRCSSP with independent and identically
distributed (iid) innovations, which allows only conditional mean dynamics, and
the pure ARCH process, which allows only conditional variance dynamics. We
then combine these extremes to produce a generalized model that permits varia-
tion in both the first and second conditional moments. Finally, we introduce the
generalized ARCH (GARCH) process, which is very useful in practice.
= (Ji LN, which are both time invariant. However, the conditional mean is time
;=0
varying and is given by E[y, I Dr-d = ~)ie,-;, where the information set is
;=1
variance 0'; ~)i. Note that for any k, the conditional prediction error variance
;=0
depends only on k and not on .a,_I; thus, readily available and potentially useful
information is discarded.
By way of contrast, we now introduce a pure ARCH process, which displays only
conditional variance dynamics. We write
Y, = E"
E, I .a,_I - N(O, h,),
h, = Cd + y(L)E;,
y, = B(L)e"
subject to the conditions discussed earlier. Both the unconditional mean and vari-
ance are constant-that is, E[y,) = 0 and
Thus, this model treats the conditional mean and variance dynamics in a symmetric
fashion by allowing for movement in each, a common characteristic of economic
time series.
y, = E"
E, I il,_, - N(O, h,),
p q
ARCH model:
h = + a(L) e2 = + ~ oe 2 .
tt /
(0 (0
I I - fJ(l) I - fJ(L) I I - ,8(1) 1-/'
The first two unconditional moments of the pure GARCH model are constant
and given by E[y,] = 0 and
GARCH models have been used extensively in macroeconomics and finance be-
cause of their attractive approximation-theoretic properties. However, these models
do not arise directly from economic theory, and various efforts have been made
to imbue them with economic rationale. Here, we discuss both approximation-
theoretic and economic motivations for the GARCH framework.
Approximation-Theoretic Considerations
The primary and most powerful justification for the GARCH model is
approximation-theoretic. That is, the GARCH model provides a flexible and par-
simonious approximation to conditional variance dynamics, in exactly the same
way that ARMA models provide a flexible and parsimonious approximation to
conditional mean dynamics. In each case, an infinite-ordered distributed lag is
approximated as the ratio oftwo finite, low-ordered lag operator polynomials. The
power and usefulness of ARMA and GARCH models come entirely from the fact
ARCH MODELS 435
that ratios of such lag operator polynomials can accurately approximate a variety
of infinite-ordered lag operator polynomials. 6 In short, ARMA models with GARCH
innovations offer a natural, parsimonious, and flexible way to capture the condi-
tional mean and variance dynamics observed in a time series.
Economic Considerations
Economic considerations may also lead to GARCH effects, although the precise
links have proved difficult to establish. Any of the myriad economic forces that
produce persistence in economic dynamics may be responsible for the appearance
ofGARCH effects in volatility. In such cases, the persistence happens to be in the
conditional second moment rather than the first.
To take one example, conditional heteroskedasticity may arise in situations in
which "economic time" and "calendar time" fail to move together. A well-known
example from financial economics is the subordinated stochastic process model of
Clark (1973). In this model and its subsequent extensions, the number of trades
occurring per unit of calendar time (I,) is a random variable, and the price change
per unit of calendar time (E,) is the sum of the I, intraperiod price changes (<5;),
which are assumed to be normally distributed:
I, i.i.d.
£, = L,<5;, <5; - N(O, 1]).
;=1
1/2 i.i.d.
E, = (1J I,) z" z, - N(O, I).
Thus, E, is characterized by conditional heteroskedasticity linked to trading vol-
ume. If the number of trades per unit of calendar time displays serial correlation,
as in Gallant, Hsieh, and Tauchen (1991), the serial correlation induced in the con-
ditional variance of returns (measured in calendar time) results in GARCH-like
behavior. Similar ideas arise in macroeconomics. The divergence between eco-
nomic time and calendar time accords with the tradition of "phase-averaging" (see
Friedman and Schwartz, 1963) and is captured by the time-deformation models of
Stock (1987, 1988).
Several other explanations for the existence of GARCH effects have been
advanced, including parameter variation (Tsay, 1987), differences in the inter-
pretability of information (Diebold and Nerlove, 1989), market microstructure
(Bollerslev and Domowitz, 1991), and agents' "slow" adaptation to news (Brock and
LeBaron, 1993). Currently, a consensus economic model producing persistence in
436 MACROECONOMETRICS
05
conditional volatility does not exist, but it would be foolish to deny the existence
of such persistence; measurement is simply ahead of theory.
(12 = co _
.' I - a(l) - 13(1)
ARCH MODELS 437
.
o.
25
20
"'~-~-_- ......--,u-----.J.
Figure 11.4. Conditional variance of the GARCH(l, 1) realization and its sample
autocorrelation function
120
100
60
40
because z; IDt-I - XZI)' However, because the median ofaXZI) is .455, p(e; < ~h,)
> 1/2. Thus, the e;
proxy introduces a potentially significant error into the analysis
of small samples of h" t = 1, ... , T, although the error diminishes as T increases.
ARCH MODELS 439
In the limit, this conditional variance reduces to the unconditional variance of the
process,
limE[el+k l.a,l = (0
k.... oo I - a(l) - 13(1)
For finite k, the dependence of the prediction error variance on the current
information set .a, can be exploited to produce better interval forecasts, as illus-
trated in Figure 11.6 for k = 1. We plot the one-step-ahead 90 percent conditional
and unconditional interval forecasts of our simulated GARCH( I, I) process along
with the actual realization. We construct the conditional prediction intervals using
the conditional variance
10
-10
o 100 200 300 400 500
Time
process with unconditional moments of all orders occurs when a(L) = f3(L) 0, =
which is the degenerate case of iid innovations. Otherwise, depending on the
precise parameterization, unconditional moments will cease to exist beyond some
point. For example, most parameter estimates for financial data indicate an infinite
fourth moment, and some even indicate an infinite second moment. Our illustra-
tive process has population mean 0, variance 10, skewness 0, and kurtosis 5.2.
As always, the likelihood function is simply the joint density of the observations,
L(8; YI, ... ,YT) = f(YI' ... ,YT; 0).
This joint density is non-Gaussian and does not have a known closed-form expres-
sion, but it can be factored into the product of conditional densities,
L(O; YI' ... ,YT) =f(YT I QT-I; O)f(YT-1 I !2T-2; 0) ...
f(Yp+1 I Qp; fJ)f(yp, ... , Yl; 0),
442 MACROECONOMETRlCS
r a . 8)
f( Y, I Ut-I, = ....r;:;:::h,()
1 - - -Yl]
8 -112 exp[1 - .
2,. 2 h,(8)
The f(y p , ••• , Y.; 8) term is often ignored because a closed-form expression for
it does not exist and because its deletion is asymptotically inconsequential. Thus,
the approximate log likelihood is
T - piT 1 T yl
In L(8; Yp+l' ... ,Yr) = - - - In(2tr) -
2
- LJn h,(8) - -
2,=p+1
L -h,(8)
2,=p+1
.
It may be maximized numerically using iterative procedures and is easily gener-
alized to models richer than the pure univariate GARCH process, such as regres-
sion models with GARCH disturbances. In that case, the likelihood is the same
with e, == Y, - E(y, l.Qt-I; 8] in place of Y,. The unobserved conditional variances
{h,(8)}?:'P+1 that enter the likelihood function are calculated at iteration j using
8 U -I), the estimated parameter vector at iterationj - I. The necessary initial values
of the conditional variance are set at the first iteration to the sample variance of
the observed data and at all subsequent iterations to the sample variance of a
simulated realization with parameters (JU-l).
The assumption of conditional normality is not always appropriate. Nevertheless,
Weiss (1986) and Bollerslev and Wooldridge (1992) show that even when normal-
ity is inappropriately assumed, the resulting quasi-MLE estimates are asymptoti-
cally normally distributed and consistent if the conditional mean and variance
functions are specified correctly. Bollerslev and Wooldridge (1992), moreover,
derive asymptotic standard errors for the quasi-MLE estimates that are robust to
conditional nonnormality and are easily calculated as functions of the estimated
parameters and the first derivatives ofthe conditional mean and variance functions.
and La; < 1. The conditional normality assumption is adopted only because it is
;=1
ARCH MODELS 443
the most common; alternative distributions can be used with no change in the
procedure. Let () = «(0, a., ... , ap).
The initial likelihood termf(yp , ••• ,YI; ()) for any given parameter configura-
tion () is simply the unconditional density of the first p observations evaluated at
(YP' ••• ,Y.l, which can be estimated to any desired degree of accuracy using well-
known techniques of simulation and consistent nonparametric density estimation.
At any iteration j, a current "best guess" of the parameter vector (ji) exists.
Therefore, a very long realization of the process with parameter vector (ji) can be
simulated and the value of the joint unconditional density evaluated at (YP' ••• , Y.l
can be consistently estimated and denoted as j(yp, ... ,Y.; (ji»). This estimated
unconditional density can then be substituted into the likelihood where the true
unconditional density appears. By simulating a large sample, the difference be-
tween!(yp , ••• 'YI; ()(j)) andf(yp , ••• ,Y.; ()(j)) is made arbitrarily small, given the
consistency of the density estimation technique. The full conditionally Gaussian
likelihood, evaluated at (fD, is then
L«()(j); YT,' .. , Y.) '" !(YP' ••• , Y.; (ji») x
IT [_l_h
T
t=P+'..fiii t
«()(j»)-1I2 ex [_ I
P
y~
2 h t«()(jJ)
]]
,
Testing
Thus, less fonnal diagnostics are often used, such as the sample autocorrelation
function of squared residuals. McLeod and Li (1983) show that under the null
hypothesis of no nonlinear dependence among the residuals from an ARMA model,
the vector of nonnalized sample autocorrelations of the squared residuals,
T
""(A2 _ A2)(A2 _ A2)
~ E, Ci EI_~ Ci
-!TPe,(r) = -!T.c::t=:.:..~+,-,-l----:rT------
L(e; - 6 2 )2
1=1
m )2
+ 2)L PC2 r
A (
Qe,(m) = T(T ,
~=I T - r
is asymptotically X~m) under the null. If the null is rejected, then nonlinear depend-
ence, such as GARCH, may be present. 14
After fitting a GARCH model, it is often of interest to test the null hypothesis
that the standardized residuals are conditionally homoskedastic. Bollerslev and
Mikkelsen (1993) argue that one may use the Ljung-Box statistic on the squared
standardized residual autocorrelations, but that the significance of the statistic
should be tested using a X~m-k) distribution, where k is the number of estimated
GARCH parameters. This adjustment is necessary due to the deflation associated
with fitting the conditional variance model.
A related testing issue concerns the effect of GARCH innovations on tests for
other deviations from classical behavior. Diebold (1987, 1988) examines the impact
of GARCH effects on two standard serial correlation diagnostiCS, the Bartlett
standard errors and the Ljung-Box statistic. As is well-known, in the large-sample
Gaussian white-noise case,
i.i.d. [ 1 ]
p(r) - N 0,
A
T
,r = 1,2, ...
and
~ 1 a
A
Q(m) = T(T+2)~ p(r)
A 2
-X(m),
2
~=I (T - T)
where p('r) denotes the sample autocorrelation at lag T. In the GARCH case,
however, an adjustment must be made,
ARCH MODELS 445
p(r)i.i.d.
A
- N (0 I[
'T I
Yy 2(r)])
+ ~ ,r = 2
I, , ... ,
y;
where Yy2 (r) denotes the autocovariance function of at lag rand (14 is the
squared unconditional variance of Yt. The adjustment is largest for small r and
decreases monotonically as r ~ 00 if the process is covariance stationary. Simi-
larly, the robust Ljung-Box statistic is
I T A
- - Ig,«(JGMM) = O.
T- m,=m+1
Note that, because there are as many parameters being estimated as there are
orthogonality conditions, GMM simply yields the standard point estimates of the
autocovariances. Their standard errors and related test statistics are asymptotically
robust, because as shown by Hansen (1982) under general conditions allowing for
heteroskedasticity and serial correlation of unknown form, .fi(8GMM - (J) aN(O, V)
where
v= (e[ag'~•• )]S-'E[ag'~GU·)Jr
and S is the spectral density matrix of gt«(J) at frequency zero. This expression
for V is made operational by replacing all population objects with consistent
446 MACROECONOMETRICS
There are numerous applications and extensions of the basic GARCH model. In
this section, we highlight those that we judge most important in macroeconomic
and financial contexts. It is natural to discuss applications and extensions simul-
taneously because many of the extensions are motivated by applications.
Numerous alternative functional forms for the conditional variance have been
suggested in the literature: 6 One of the most interesting is Nelson's (1991) expo-
nential GARCH(p, q) or EGARCH(p, q) model,
i.i.d.
Z, - N(O, I),
p q
Consider a regression model with GARCH disturbances of the usual sort, with one
additional twist: the conditional variance enters as a regressor, thereby affecting
the conditional mean. Write the model as
A special case of the GARCH model is the integrated GARCH (IGARCH) model,
introduced by Engle and Bollerslev (1986). A GARCH(p, q) process is integrated
of order one in variance if 1 - a(L) - f3(L) = 0 has a root on the unit circle. The
IGARCH process is potentially important because, as an empirical matter, GARCH
roots near unity are common in high-frequency financial data.
The earlier ARMA result for the squared GARCH process now becomes an
ARIMA result for the squared IGARCH process. As before, e;
= (0 + [a(L) +
f3(L)]E; - f3(L)v, + v,; thus, [I - a(L) - f3(L)]E; = (0 - f3(L)v, + v,, When the
autoregressive polynomial contains a unit root, it can be rewritten as
[I - a(L) - f3(L)]E; = q>(L)(I-L)e; = (0 - f3(L)v, + v,,
Thus, the differenced squared process is of stationary ARMA form.
Unlike the conditional prediction error variance for the covariance stationary
GARCH process, the IGARCH conditional prediction error variance does not
converge as the forecast horizon lengthens; instead, it grows linearly with the
length of the forecast horizon. Formally, E[E;+k I Q,] = (k - 1)(0 + ht+1 so that
lim E[ E~+k I .Q I] = 00. Thus, the IGARCH process has an infinite unconditional
k....-
variance.
Clearly, a parallel exists between the IGARCH process and the vast litera-
ture on unit roots in conditional mean dynamics (see Stock, 1994). This parallel,
448 MACROECONOMETRICS
£1 =crlZ I = ex p[ ~ }I'
Z, - N(O, 1),
h, = ro + (3hH + TI"
TIl - N(O, cr~).
Thus, as opposed to standard GARCH models, hI is not deterministic conditional
on Qt-l; the conditional variance evolves as a first-order autoregressive process
driven by a separate innovation. Moreover, the exponential specification ensures
that the conditional variance remains positive. It is clear that the stochastic volatility
ARCH MODELS 449
q p
vech(H, ) = W + LAvech(e,_;e,~;) + LB;vech(H,),
;=( j=(
vech(.) is the vector-half operator that converts (Nx N) matrices into (N(N + 1)/2
x I) vectors of their lower triangular elements, W is an (N(N + 1)/2 x I) parameter
vector, and Ai and Bj are «N(N + 1)/2) x (N(N + 1)/2)) parameter matrices.
Likelihood-based estimation and inference are conceptually straightforward and
parallel the univariate case. The approximate log likelihood function for the con-
ditionally Gaussian multivariate GARCH(p, q) process, aside from a constant, is
I TIT
In L(O; ej , . . . , ET) =- - LInIH,1 - - Le', H,-I E, ; j ~ l.
2,=j 2,=j
Multivariate models with factor structure, such as the latent-factor GARCH model
(Diebold and Nerlove, 1989) and the factor GARCH model (Engle, 1987; Bollerslev
and Engle, 1993), capture the idea of commonality of volatility shocks, which
appears empirically relevant in systems of asset returns in the stock, foreign
exchange, and bond markets. 18 Models with factor structure are also parsimonious
and are easily constrained to maintain positive definiteness of the conditional
covariance matrix.
In the latent-factor model, movements in each of the N time series are driven
by an idiosyncratic shock and a set of k < N common latent shocks or "factors.,,19
The latent factors display GARCH effects, whereas the idiosyncratic shocks are
i.i.d. and orthogonal at all leads and lags. The one-factor model is important in
practice, and we describe it in some detail. The model is written as £, = AF,+ v"
where £" A, and v, are (N x 1) vectors and F, is a scalar. F, and v, have zero
conditional means and are orthogonal at all leads and lags. The factor F, follows
a GARCH(p, q) process,
F, 1[2'_1 - N(O, hI)
h, = OJ + a(L)F; + f3(L)h"
so that the conditional distribution of the observed vector is
£, I [2'_1 - N(O, H,),
H, = U'h, + r,
where r= cov(v,) = diag(YI, ... , rN). Thus, thej'h time-t conditional variance is
TIme
Hi'" h,
= A/., =J./.,['" + t.a,F,~ + tp,h,-l
Note that the latent factor F, is unobservable and not directly included in D,_, =
{E'_I' ... , E(}. Effectively, the latent-factor model is a stochastic volatility model.
In general, the number of parameters in the k-factor model is N(k + I) + K(1
+ p + q) = O(N), so the number of parameters in the one-factor case is 2N + (1
+ P + q), a drastic reduction relative to the general multivariate case. Moreover,
the conditional covariance matrix is guaranteed to be positive definite, so long as
the conditional variances of the common and idiosyncratic factors are constrained
to be positive.
A simulated realization from a bivariate model with one common GARCH(I,
1) factor is shown in Figures 11.7, 11.8, and 11.9. The model is parameterized as
[El'] = [0.6]
E21 0.9
F, + [VI'],
V21
h, = I + .2F;_1 + .7h'_I'
(VI', V2,)'
i.i.d.
- N(O, I).
The realization of the common factor underlying the system is precisely the one
presented in our earlier discussion of univariate GARCH models. The latent-factor
GARCH series exhibit the volatility clustering present in the common factor. As
452 MACROECONOMETRICS
eo eo
70 70
eo 60
'" '"
40 40
30
T1mo
before, the squared realizations of the two series indicate a degree of persistence
in volatility. Furthennore, as expected, the conditional second moments of the two
series are similar to that of F, because, as shown above, they are simply multiples
of h,.
Diebold and Nerlove (1989) suggest a two-step estimation procedure. The first
step entails performing a standard factor analysis-that is, factoring the uncondi-
tional covariance matrix as H = AX (12 + r, where (12 is the unconditional variance
of F" and extracting an estimate of the time series of factor values {F;}:~J. The
second step entails estimating the latent-factor GARCH model treating the ex-
tracted factor series F; as if it were the actual series F,.
The Diebold-Nerlove procedure is clearly suboptimal relative to fully simul-
taneous maximum likelihood estimation because the F, series is not equal to the F,
series, even asymptotically. Harvey, Ruiz, and Sentana (1992) provide a better
approximation to the exact likelihood function that involves a correction factor to
account for the fact that the F, series is unobservable. 20 For example, using an
ARCH( 1) specification, the conditional variance of the latent factor F, in the
Diebold-Nerlove model is
30
25 05
20
.5
30
25 05
20
-o.,A-----r---nr-----T>~- __-_______.I.
T..,.
30
25 05
20
'5
-o.i\-----r--_n.__----nc---...._-_,/.
T..,.
Figure 11.9. The conditional variance of the two factor-GARCH series and
their conditional covariance as well on the corresponding sample autocorrelation
functions
454 MACROECONOMETIUCS
Volatility forecasts are readily generated from GARCH models and used for a
variety of purposes, such as producing improved interval forecasts, as discussed
previously. Less obvious but equally true is the fact that, under asymmetric loss,
volatility dynamics can be exploited to produce improved point forecasts, as shown
by Christoffersen and Diebold (1994). If, for example, Y'+kis normally distributed
with conditional mean Jl'+k I il, and conditional variance h'+kl il, and L(e'+k) is any
loss function defined on the k-step-ahead prediction error e'+k = Yr+k - Yl+k' then the
optimal predictor is Y'+k = J.l'+k I il, + a" where a, depends only on the loss
function and the conditional prediction error variance var(e'+k I il,) = var(Y'+k I ilJ
= h'+k I il,. The optimal predictor under asymmetric loss is not the conditional
mean, but rather the conditional mean shifted by a time-varying adjustment that
depends on the conditional variance. The intuition for this is simple: when, for
example, positive prediction errors are more costly than negative errors, a negative
conditionally expected error is desirable and is induced by setting the bias a, > O.
The optimal amount of bias depends on the conditional prediction error variance
of the process. As the conditional variation around J.l t+k I il, grows, so too does the
optimal amount of bias needed to avoid large positive prediction errors.
To illustrate this idea, consider the linlin loss function, so-named for its linear-
ity on each side of the origin (albeit with possibly different slopes):
L(Yt+k _ Y'+k =
A ) {ab II Y'+k - Y,+k I'f
I. if YI+k - Yt+k > 0
<0
Yt+k - Yt+k
A
,1 Y,H - Yt+k -
A
•
Christoffersen and Diebold (1994) show that the optimal predictor of Y,+k under
this loss function is
ARCH MODELS 455
-150k------:1;-;!;O=O-------..2~OO.-------..,3~O""O------,4ckOO,,---F,500
Although volatility forecast accuracy comparisons are often conducted using mean-
squared error, loss functions that explicitly incorporate the forecast user's economic
loss function are more relevant and may lead to different rankings of models.
West, Edison, and Cho (1993) and Engle et al. (1993) make important contributions
along those lines, proposing economic loss functions based on utility maximization
and profit maximization, respectively.
Lopez (1995) proposes a volatility forecast evaluation framework that sub-
sumes a variety of economic loss functions. The framework is based on transform-
ing a model's volatility forecasts into probability forecasts by integrating over the
distribution of c,. By selecting the range of integration corresponding to an event
of interest, a forecast user can incorporate elements of her loss function into the
probability forecasts. For example, given c, I .Q,_I - D(O, hI) and a volatility
forecast hr , an options trader interested in the event c, E [Lc." Uc.,] would generate
the probability forecast
where iL, is the forecasted conditional mean and (ly.T+t, U.y,T+') is the standardized
range of integration.
The probability forecasts so-generated can be evaluated using statistical tools
tailored to the user's loss function, In particular, probability scoring rules can be
used to assess the accuracy of the probability forecasts, and the significance of
differences across models can be tested using a generalization of the Diebold-
Mariano (1995) procedure, Moreover, the calibration tests of Seillier-Moiseiwitsch
and Dawid (1993) can be used to examine the degree of equivalence between an
ARCH MODELS 457
Fifteen years ago, little attention was paid to conditional volatility dynamics in
modeling macroeconomic and financial time series; the situation has since changed
dramatically. GARCH and related models have proved tremendously useful in
modeling such dynamics. However, perhaps in contrast to the impression we may
have created, we believe that the literature on modeling conditional volatility
dynamics is far from settled and that complacency with the ubiquitous GARCH( I,
1) model is not justified.
Almost without exception, low-ordered (and hence potentially restrictive)
GARCH models are used in applied work. For example, among hundreds of
empirical applications of the GARCH model, almost all casually and uncritically
adopt the GARCH( I, 1) specification. EGARCH applications have followed suit
with the vast majority adopting the EGARCH(l, 1) specification. Similarly, appli-
cations of the stochastic volatility model typically use an AR( 1) specification.
However, recent findings suggest that such specifications-as well as the models
themselves, regardless of the particular specification-are often too restrictive to
maintain fidelity to the data.
It appears, for example, that the conditional volatility dynamics of stock market
returns (as well as certain other asset returns) contain long memory. Ding, Engle,
and Granger (1993) find positive and significant sample autocorrelations for daily
S&P 500 returns at up to 2,500 lags and that their rate of decay is slower than
exponential. A model consistent with such long-memory volatility findings is
the fractionally integrated GARCH (FIGARCH) model developed by Baillie,
Bollerslev, and Mikkelsen (1993), building on earlier work by Robinson (1991).
FIGARCH is a model of fractionally integrated conditional variance dynamics, in
parallel to the well-known fractionally integrated ARMA models of conditional
mean dynamics (see Granger and Joyeux, 1980). The FIGARCH model implies
a hyperbolic rate of decay for the autocorrelations of the squared process that is
slower than exponential.
To motivate the FIGARCH process, begin with the GARCH(l, 1) process,
h, = CO + a(L)E~ + f3(L)h,.
458 MACROECONO~CS
.
: "'--~~~---.......----.,..~
------========-=======
Figure 11.11. Autocorrelation function-SaP ( lei Jan. 1928 to May 1990). and
autocorrelation function-SaP (e 2 , Jan. 1928 to May 1990)
Rearranging the conditional variance into ARMA fonn, the FIGARCH (p, d, q)
equation is
[I - a(L) - f3(L) ]e; = 4'(L)(1 - L)de; = (J) + (I - f3(L) )v,.
That is, the [1 - a(L) - f3(L)] polynomial can be factored into a stationary ARMA
component and a long-memory difference operator. If 0 < d < 1, the process is
FIGARCH(p, d, q). If d = 0, then the standard GARCH(p, q) model obtains; if d
= 1, then the IGARCH(p, q) model obtains. Bollerslev and Mikkelsen (1993)
conjecture that the coefficients in the ARCH representation of a FIGARCH pro-
cess (d < 1) are dominated by those of an IGARCH process. If so, then FIGARCH
(d < 1) would be strictly stationary (though not covariance stationary) because
IGARCH is strictly stationary.
Long memory is only one of many previously unnoticed features of volatility.
Interestingly, as we study volatility more carefully, more and more anomalies
emerge. Volatility patterns tum out to differ across assets, time periods, and trans-
formations of the data. The complacency with the "standard" CARCH model is
being shattered, and we think it unlikely that anyone consensus model will take
its place. The implications of this development are twofold. First, real care must
be taken in tailoring volatility models to the relevant data, as in Engle and Ng
(1993). Second, because all volatility models are likely to be misspecified, care
should be taken in assessing models' robustness to misspecification.
To illustrate the deviations from classical GARCH models that tum out to be
routinely present in real data, we present in Figure 11.11 the sample autocorrelation
functions of the absolute and squared change in the log daily closing value of the
S&P 500 stock index, 1928-1990. The autocorrelation functions are shown to
displacement 't'= 200 in order to assess the evidence for long memory, and dashed
lines indicate the Bartlett 95 percent confidence interval for white noise. Note that
ARCH MODELS 459
A I ;L _
Figure 11.12. Autocorrelation function-S&P (£2, Jan. 1928 to Dec. 1940), auto-
correlation function-S&P(£2, Jan. 1941 to Dec. 1970), autocorrelation function-
S&P (£2, Jan. 1971 to Dec. 1980), and autocorrelation function-S&P (£2, Jan.
1981 to May 1990)
often seem to indicate structural change. For example, the long memory seemingly
present in exchange rate volatility seems concentrated in the 1970s, while long
memory in interest rate volatility is typically concentrated in the 1980s. These
observed phenomena, as well as occasional long-horizon spikes in autocorrelations
and the appearance of oscillatory autocorrelation behavior, are again inconsistent
with standard specifications.
An additional illustration of the inadequacies of GARCH models is provided
by West and Cho (1994). Using weekly exchange rates, they show that for hori-
zons longer than one week, out-of-sample GARCH volatility forecasts loose their
value, even though volatility seems highly persistent. The good in-sample per-
formance of GARCH models breaks down rapidly out-of-sample. 21 In addition,
standard tests of forecast optimality, such as regressions of realized squared returns
on an intercept and the GARCH forecast, strongly reject the null of the optimality
of the GARCH forecast with respect to available information. West and Cho
(1994) suggest time-varying parameters and discrete shifts in the mean level of
volatility as possible explanations.
In light of the emerging evidence that GARCH models are likely misspecified
and the unlikely occurrence of happening upon a "correct" specification, it is of
interest to consider whether GARCH models might still perform adequately in
tracking and forecasting volatility-that is, whether their good properties are ro-
bust to misspecification. In a series of papers (Nelson, 1990a, 1992, 1993; Nelson
and Foster, 1991, 1994), Nelson and Foster find that the usefulness of GARCH
models in volatility tracking and short-term volatility forecasting is robust to a
variety of types of misspecification; thus, in spite of misspecification, GARCH
models can consistently extract conditional variances from high-frequency time
series. More specifically, if a process is well approximated by a continuous-time
diffusion, then broad classes of GARCH models provide consistent estimates of
the instantaneous conditional variance as the sampling frequency increases. This
occurs because the sequence of GARCH(I, 1) models used to form estimates of
next period's conditional variance average increasing numbers of squared residuals
from the increasingly recent past. In this way, a sequence of GARCH(l , 1) models
can consistently estimate next period's conditional variance despite potentially
severe misspecification.
Acknowledgments
Helpful comments were received from Richard Baillie, Tim Bollerslev, Pedro de
Lima, Wayne Ferson, Kevin Hoover, Peter Robinson, and Til Schuermann. We
gratefully acknowledge the support of the National Science Foundation, the Sloan
Foundation, and the University of Pennsylvania Research Foundation.
ARCH MODELS 461
Notes
21. Note, however, that West and Cho (1994) evaluate volatility forecasts using the mean-squared
error criterion, which may not be the most appropriate. For further discussion, see Bollerslev, Engle,
and Nelson (1994) and Lopez (1995).
References
Baillie, R.T., T. Bollerslev, and H.O. Mikkelsen. (1993). "Fractionally Integrated Gener-
alized Autoregressive Conditional Heteroskedasticity." Manuscript, J.L. Kellogg School
of Management, Northwestern University.
Bera, A.K., and M.L. Higgins. (1993). "ARCH Models: Properties, Estimation and Test-
ing." Journal of Economic Surveys 7, 305-362.
Black, F. (1976). "Studies of Stock Price Volatility Changes." Proceedings ofthe American
Statistical Association, Business and Economic Statistic Section (pp. 177-181).
Bollerslev, T. (1986). "Generalized Autoregressive Conditional Heteroskedasticity." Jour-
nal of Econometrics 31,307-327.
Bollerslev, T. (1987). "A Conditional Heteroskedastic Time Series Model for Speculative
Prices and Rates of Return." Review of Economics and Statistics 69,542-547.
Bollerslev, T. (1988). "On the Correlation Structure for the Generalized Autoregressive
Conditional Heteroskedastic Process." Journal of Time Series Analysis 9, 121-131.
Bollerslev, T., R.Y. Chou, and K.F. Kroner. (1992). "ARCH Modeling in Finance: A
Selective Review of the Theory and Empirical Evidence." Journal ofEconometrics 52,
5-59.
Bollerslev, T., and I. Domowitz. (1991). "Price Volatility, Spread Variability and the Role
of Alternative Market Mechanisms." Review of Futures Markets 10,78-102.
Bollerslev, T., and R.F. Engle. (1993). "Common Persistence in Conditional Variances."
Econometrica 61, 166-187.
Bollerslev, T., R.F. Engle, and D.B. Nelson. (1994). "ARCH Models." In R.F. Engle and
D. McFadden (eds.), Handbook of Econometrics (Vol. 4). Amsterdam: North-Holland.
Bollerslev, T., R.F. Engle, and J.M. Wooldridge. (1988). "A Capital Asset Pricing Model
with Time Varying Covariances." Journal of Political Economy 95, 116-131.
Bollerslev, T., and H.O. Mikkelsen. (1993). "Modeling and Pricing Long Memory in Stock
Market Volatility." Manuscript, J.L. Kellogg School of Management, Northwestern
University.
Bollerslev, T., and J.M. Wooldridge. (1992). "Quasi-Maximum Likelihood Estimation and
Inference in Dynamic Models with Time-Varying Covariances." Econometric Reviws
II, 143-179.
Box, G.E.P., and G.W. Jenkins. (1970). Time Series Analysis Forecasting and Control.
Oakland: Holden-Day.
Brock, W.A., and B.D. LeBaron. (1993). "Using Structural Modeling in Building Statistical
Models of Volatility and Volume of Stock Market Returns." Manuscript, Department
of Economics, University of Wisconsin, Madison.
Christoffersen, P.F., and F.x. Diebold. (1994). "Optimal Prediction under Asymmetric
Loss." Technical Working Paper 167, National Bureau of Economic Research.
ARCH MODELS 463
Friedman, M., and AJ. Schwartz. (1963). A Monetary History of the United States, 1867-
1960. Princeton, NJ: Princeton University Press.
Gallant, A.R., D.A. Hsieh, and G. Tauchen. (1991). "On Fitting a Recalcitrant Series: The
Pound-Dollar Exchange Rate, 1974-1983." In W.A. Barnett, J. Powell, and G. Tauchen
(eds.), Nonparametric and Semiparametric Methods in Econometrics and Statistics.
Cambridge: Cambridge University Press.
Geweke, J. (1989). "Bayesian Inference in Econometric Models Using Monte Carlo Inte-
gration." Econometrica 57, 1317-1339.
Granger, C.WJ., and R. Joyeux. (1980). "An Introduction to Long-Memory Time Series
Models and Fractional Differencing." Journal of Time Series Analysis 1, 15-39.
Granger, C.WJ., and P. Newbold. (1979). Forecasting Economic Time Series. New York:
Academic Press.
Hamilton, J.D., and R. Susmel. (1994). "Autoregressive Conditional Heteroskedasticity
and Changes in Regime." Journal of Econometrics 64,307-333.
Hansen, L.P. (1982). "Large Sample Properties of the Method of Moment Estimators."
Econometrica 50, 1029-1054.
Harvey, A., E. Ruiz, and E. Sentana. (1992). "Unobserved Component Time Series Models
with ARCH Disturbances." Journal of Econometrics 52, 129-158.
Harvey, A., E. Ruiz, and N. Shephard. (1994). "Multivariate Stochastic Variance Models."
Review of Economic Studies 61, 247-264.
Jacquier, E., N.G. Polson, and P.E. Rossi. (1994). "Bayesian Analysis of Stochastic Vola-
tility Models." Journal of Business and Economics Statistics 12, 371-389.
Jorgenson, D.W. (1966). "Rational Distributed Lag Functions." Econometrica 34,135-149.
Kim, S., and N. Shephard. (1994). "Stochastic Volatility: Likelihood Inference and Com-
parison with ARCH Models." Manuscript, Nuffield College, Oxford University.
King, M., E. Sentana, and S. Wadhwani. (1994). "Volatility and Links Between National
Stock Markets." Econometrica 62, 901-933.
Kraft, D., and R.F. Engle. (1982). "Autoregressive Conditional Heteroskedasticity in Multiple
Time Series Models." Discussion Paper 82-23, Department of Economics, University
of California, San Diego.
Lamoureux, C.G., and W.D. Lastrapes. (1990). "Persistence in Variance, Structural Change
and the GARCH Model." Journal of Business and Economic Statistics 8, 225-234.
Lastrapes, W.D. (1989). "Exchange Rate Volatility and U.S. Monetary Policy: An ARCH
Application." Journal of Money, Credit and Banking 21, 66-77.
Lee, J.H.H. (1991). "A Lagrange Multiplier Test for GARCH Models." Economics Letters
37, 265-271.
Lee, J.H.H., and M.L. King. (1993). "A Locally Most Mean Powerful Based Score Test for
ARCH and GARCH Regression Disturbances." Journal of Business and Economics
Statistics 11, 17-27.
Lee, S.-W., and RE. Hansen. (1994). "Asymptotic Theory for the GARCH (I, I) Quasi-
Maximum Likelihood Estimator." Econometric Theory 10, 29-52.
Lopez, J.A. (1995). "Evaluating the Predictive Accuracy of Volatility Models." Manu-
script, Department of Economics, University of Pennsylvania.
Lumsdaine, R.L. (1992). "Asymptotic Properties of the Quasi-Maximum Likelihood
ARCH MODELS 465
Tsay, R.S. (1987). "Conditional Heteroskedastic Time Series Models." Journal of the
American Statistical Association 82, 509-604.
Weiss, A.A. (1984). "ARMA Models with ARCH Errors." Journal of Time Series Analysis
5, 129-143.
Weiss, A.A. (1986). "Asymptotic Theory for ARCH Models: Estimation and Testing."
Econometric Theory 2, 107-131.
West, K.D., and D. Cho (1994). 'The Predictive Ability of Several Models of Exchange
Rate Volatility." Technical Working Paper 152, National Bureau of Economic Research.
West, K.D., H.J. Edison and D. Cho (1993). "A Utility-Based Comparison of Some Models
of Exchange Rate Volatility." Journal of International Economics 35, 23-45.
White, H. (1984). Asymptotic Theory for Econometricians. New York: Academic Press.
Wold, H.O. (1938). The Analysis ofStationary Time Series. Uppsala: Almquist and Wicksell.
Commentary on Chapter 11
Douglas G. Steigerwald
Frank Diebold and Jose Lopez have written an excellent primer on conditional
heteroskedasticity (CH) models and their use in applied work. A principal moti-
vation for CH models, as outlined in Diebold and Lopez, is their ability to parsi-
moniously capture the observed characteristics of many financial time series. By
far the most widely used CH model, in part because of the fact that estimators for
the model are simple to construct, is the generalized autoregressive conditional
heteroskedasticity (GARCH) specification of order (1, I) with normal innovations
(henceforth termed the normal GARCH (1, I) model). Despite its widespread
use, the normal GARCH(l, 1) model does not account for important features in
many financial time series. For example, assuming that the GARCH innovations
have a normal density generates far fewer outliers than are typically observed
in asset prices, while assuming that the order of the GARCH model is (1, 1) fails
to account for the variety of dynamic patterns observed in the conditional
heteroskedasticity of asset prices.
As Diebold and Lopez note in describing avenues for future research, it is
important to consider alternative CH models that do account for such features of
asset prices. Two alternatives to a normal GARCH(1, 1) model, which are men-
tioned by Diebold and Lopez and for which estimators are also simple to con-
struct, are (1) to allow for nonnormal innovations that have a thicker tailed density,
thereby accounting for a larger number of outliers and (2) to allow for orders other
than (1, 1) by developing powerful test statistics for selection of order in GARCH
models, thereby accounting for a wider variety of dynamic patterns. I discuss each
of these alternatives in tum, in an effort to bring them within the set of commonly
used methods for estimation and testing of CH models.
Unknown Density
Let y, be a period-t variable (such as an exchange rate) that has conditional mean
x,fJ where x,£9t k and the period-t regressors include a constant. The normal
GARCH(1, 1) model for y, is
(1)
where the period-t conditional variance is
467
468 MACROECONOMETIUCS
(2)
with period-t innovation u, and where (/3', m, aI' y.) are parameters to be estim-
ated. The sequence {u, };=I is a sequence of independent and identically distributed
(iid) normal random variables with mean zero and variance one.!
Because the normal GARCH(I, I) model does not adequately account for
outliers in asset prices such as exchange rates, researchers constructing CH models
of exchange rates often assume that the density of u, has thicker tails than a normal
density. For example, Baillie and Bollerslev (1989) use both an exponential-power
and a t density to model exchange rates.
Although the use of thicker-tailed parametric innovation densities does account
for a larger number of outliers, it also raises the issue of the properties of the
estimators if the selected density is misspecified. Virtually all researchers that
estimate CH models also use a quasi-maximum likelihood estimator (QMLE). If
the assumed density is normal, the QMLE is consistent for the parameters of the
conditional variance. If the assumed density is nonnormal, then consistency of the
QMLE depends on the specification of the conditional mean. For a nonnormal
GARCH(I, I) model, which is given by (I) and (2) together with the assumption
that u, has a nonnormal density, Newey and Steigerwald (1994) show that a non-
normal QMLE is not generally consistent. 2
Analternative estimator that also accounts for a larger number of outliers is a
semiparametric estimator. A semiparametric estimator of the parameters in a
GARCH( I, I) model is constructed under the assumption that the innovation
density is any member within a class of densities, and uses a nonparametric estim-
ator of the density. Steigerwald (1994) shows that a semiparametric estimator is
consistent for general GARCH(p, q) models.
Given that a semiparametric estimator accounts for a larger number of outliers
and consistently estimates the parameters of (I) and (2), attention turns to finite
sample performance. The finite sample performance of a semiparametric estimator
depends on the bandwidth used to construct the nonparametric density estimator.
The bandwidth, in turn, depends on the conditional variance parameterization. For
the conditional variance parameterization (2), the regularity conditions given in
Steigerwald require that the bandwidth used to construct the nonparametric den-
sity estimator be smaller than the optimal bandwidth. Such a restriction on choice
of bandwidth may lead to a poor estimate of the density, thereby reducing the
gains of a semiparametric estimator.
A reparameterization of the conditional variance, which allows the optimal
bandwidth to be used to estimate the density, is to let the variance of u, be re-
stricted only to be finite and to reparameterize the conditional variance as
(3)
COMMENTARY ON CHAPTER 11 469
Linton (1993) develops this reparameterization for ARCH models, Drost and
Klaasen (1993) and Steigerwald extend the reparameterization to GARCH
models. 3
A guide to the finite sample performance of semiparametric estimators for the
two conditional variance parameterizations is provided by the simulations con-
tained in Engle and Gonzalez-Rivera (1991) and Steigerwald. Both studies com-
pare the performance of a semiparametric estimator with a normal QMLE. For a
sample size of 2,000, Engle and Gonzalez-Rivera report essentially no gain in
efficiency for a semiparametric estimator of the parameters in (I) and (2) when the
density of ", is a t density with 5 degrees of freedom. Steigerwald, using a different
nonparametric estimator, reports more favorable results for a semiparametric
estimator of the parameters in (I) and (2), finding some efficiency gains with
a sample of only fifty observations when the density for ", is a t density with 5
degrees of freedom. The efficiency gains increase dramatically if (3) replaces (2),
indicating that the parameterization of the conditional variance is important for
applied work.
All discussion in the preceding section considers a fixed order (I, I) for the
conditional variance. As is noted in the introduction, the (I, I) order specification
fails to account for the variety of dynamic patterns in many time series. In re-
sponse, I tum to extending the GARCH(I, 1) specification to general order (p, q).
To extend the GARCH(l, I) specification to GARCH(p, q) requires a test
statistic for choosing correct order. To keep the following discussion of test sta-
tistics clear, I consider two distinct testing problems. The first problem is to test
the null hypothesis that the conditional variance is ARCH(p) against the univariate
alternative hypothesis that the conditional variance is ARCH(p + 1). The second
problem is to test the null hypothesis that the conditional variance is ARCH(p)
against the multivariate alternative hypothesis that the conditional variance is
ARCH(p + k), where k > I. In particular, the two problems can be viewed as
testing the null hypothesis of homoskedasticity against the alternative either of
ARCH(I) or ARCH(p).4
Engle (1982) develops Lagrange multiplier (LM) test statistics for the univariate
testing problem in a normal ARCH model. The test is two sided, the null hypo-
thesis that a specific conditional variance parameter equals zero is tested against
the alternative that the parameter is nonzero. Yet to ensure that the conditional
variance is always positive, the parameter of the conditional variance must be non-
negative. Therefore, more powerful test statistics can be constructed that are one
sided. For the univariate testing problem, the signed square-root of the LM test
470 MACROECONOMETIUCS
statistic provides such a one-sided test. For the multivariate testing problem, there
is no uniformly best one-sided test because the region over which the power func-
tion is evaluated spans more than one dimension. Lee and King (1993) propose a
one-sided test statistic, termed an LBS test statistic, for the multivariate testing
problem that maximizes the average slope, over all directions, of the power func-
tion in a neighborhood of the null hypothesis. They show that their one-sided test
can be more powerful in finite samples than a two-sided LM test statistic.
Both the LM test statistic and the LBS test statistic are constructed under the
assumption that the innovation density is normal. Fox (l994a) develops semi-
parametric versions of both test statistics. He finds that incorporating a nonpara-
metric estimator of the density can have important finite sample consequences.
Specifically, for samples of 100 observations the serniparametric test statistics
achieve size-adjusted power gains of as much as 20 percent over their parametric
counter-parts. Linton and Steigerwald (1994) extend the semiparametric tests to
the reparameterization of the conditional variance in (3) and show that the semi-
parametric tests are optimal in that they maximize the average slope of the power
function in a neighborhood of the null hypothesis for any innovation density in a
general class. Fox also finds that testing for correct order is important in estima-
tion, incorrect order specification can lead to substantial bias in the estimators of
the conditional variance parameters.
Empirical Implementation
notes, apparently small differences in the point estimates of the conditional vari-
ance parameters can have important economic consequences. He shows that op-
timal portfolio weights based on the estimated conditional variance process differ
markedly for two sets of estimates that differ only slightly, as do those in Table
11.1. The portfolio weights implied by the semiparametric estimators are better,
in the sense that the risk associated with a portfolio that provides a fixed expected
return is reduced by 8 to 10 percent out-of-sample.
Although the GARCH(I, 1) specification is common in the empirical finance
literature, it may not adequately account for the dynamic pattern in the data. To
test for incorrect order specification, I test the null hypothesis that the conditional
variance is GARCH(l, 1) against the alternative hypothesis that the conditional
variance is GARCH(2, 1). I construct both parametric and semiparametric versions
of the LM test statistic and the King and Lee test statistic. Both the parametric and
semiparametric LM test statistics, which are two-sided tests, fail to reject the null
hypothesis. The parametric King and Lee test statistic also fails to clearly reject
the null hypothesis. Only the semiparametric King and Lee test statistic clearly
rejects the null hypothesis. It appears that if the true dynamic process is richer than
a GARCH(l, 1), the power gains available from both a nonpararnetric density
estimator and a one-sided alternative are needed to detect it.
In summary, recent advances in econometric methodology have provided
researchers with powerful tools to move beyond the restrictive framework of a
normal GARCH(I, 1) model. Semiparametric estimators and test statistics are
available and provide alternatives that more flexibly account for the wide variety
of patterns in financial time series.
Notes
1. The variance of ", is assumed to equal one because (co. a l • r,) and the scale of ", are not
separately identified.
472 MACROECONOMETRICS
2. To ensure that the likelihood has a unique maximum, which is a necessary condition for
consistent estimation, the set of regressors must include the conditional standard deviation.
3. In (3) the parameter wcannot be separately identified because the variance of u, is restricted only
to be finite, so only ratios of the parameters (namely e"'a I and e/l)r,) are identified.
4. Considering only ARCH processes is not restrictive, as Lee and King (1993) note testing a null
hypothesis of homoscedasticity against an alternative hypothesis of ARCH( p) is equivalent to testing
against an alternative of GARCH(p, q).
References
Baillie, R., and T. Bollerslev. (1989). "The Message in Daily Exchange Rates: A
Conditional-Variance Tale." Journal of Business and Economic Statistics 7, 297-305.
Drost, F., and C. Klaasen. (1993). "Adaptivity in Semipararnetric GARCH Models." Manu-
script, Tilburg University.
Engle, R. (1982). "Autoregressive Conditional Heteroscedasticity with Estimates of the
Variance of U.K. Inflation." Econometrica 50, 987-1008.
Engle, R., and G. Gonzalez-Rivera. (1991). "Semipararnetric ARCH Models." Journal of
Business and Economic Statistics 9,345-360.
Fox, S. (1994a). "Semiparametric Testing of Generalized ARCH Processes." Manuscript,
University of California, Santa Barbara.
Fox, S. (l994b). "Hedge Estimation Using Semiparametric Methods." Manuscript, Univer-
sity of California, Santa Barbara.
Linton, 0., and D. Steigerwald. (1994). "Efficient Testing in GARCH Models." Manu-
script, Yale University.
Lee, J., and M. King. (1993). "A Locally Mean Most Powerful Score Based Test for ARCH
and GARCH Disturbances." Journal of Business and Economic Statistics II, 17-27.
Linton, O. (1993). "Adaptive Estimation in ARCH Models." Econometric Theory 9, 539-
569.
Newey, W., and D. Steigerwald. (1994). "Consistency of Quasi-Maximum Likelihood
Estimators in Models with Conditional Heteroskedasticity." Manuscript, University of
California, Santa Barbara.
Steigerwald, D. (1994). "Efficient Estimation of Financial Models with Conditional
Heteroskedasticity." Econometric Theory, forthcoming.
12 DYNAMIC SPECIFICATION
AND TESTING FOR UNIT ROOTS
AND COINTEGRATION
Anindya Banerjee
Introduction
In the field of modeling economic time series, the I980s might easily be described
as the decade of cointegration. During this decade, theoretical and applied
econometricians alike invested a great deal of effort in dealing with the theoretical
and empirical implications of Nelson and Plosser's (1982) central observation that
time series of important economic variables such as consumption and per capita
GNP may have statistical properties quite distinct from those that would warrant
the use of standard tools, such as normal, t-, and F-tables, of inference and
estimation.
The results of this research have given rise to an unmanageably vast literature
on almost every aspect of estimation and inference in the presence of nonstationary
series. It is therefore impossible, in the space available, to provide a complete
account of this field. The case for writing a formal survey is in any case rather
limited, given the several surveys that have already appeared (see, for example,
Stock and Watson, 1988; Dolado, Jenkinson, and Rivero, 1990; Campbell and
Perron, 1991). The purpose of this paper is instead to focus on specific issues that,
in my view, are important in any evaluation of this literature. In particular, I will
473
474 MACROECONOMETIUCS
attempt to address, at least partially, the issue of the extent to which our notions
of what constitutes good or appropriate modeling practice have changed as a result
of the research on unit roots.
The mathematical and statistical tools on which the econometrics literature on
unit roots depends, date back at least to the 1920s, 1930s, and 1940s, notably to
the work of Wiener, Levy, Doob, and many others. Thus no claim for math-
ematical or statistical originality can be made, per se, on behalf of this literature.
Rather, I will argue that econometricians have brought the highly developed theory
on stochastic processes into the realm of every-day econometric modeling and, by
applying it to problems particular to econometrics, have added considerably to our
store of knowledge of the very special properties of nonstationary series and the
implications of these properties for estimation and inference.
However, I will also argue that while a greater realization of how things can
go wrong when dealing with these series has had an important effect on our
econometric consciousness, this does not lead us necessarily to a fundamental
reevaluation of modeling practice, in particular of dynamic modeling. For example,
it will be shown that some of the inferential problems that arise may be overcome
by suitable transformations and appropriate augmentation of equations. These
methods will often allow us to retum to using standard tables for inference.
In the next section, two examples are presented that illustrate the fundamental
differences that could arise between the treatments of unit root and non-unit root
processes. The differences emerge particularly because the critical values of stand-
ard tests, such as t- or F- tests, are affected by the presence or absence of unit
roots. In the third section it is shown how these differences can, in some circum-
stances, be eliminated by a proper reformulation of the model. However, there are
cases where such reformulations are not possible, and it therefore becomes very
important, before proceeding to the formal task of econometric modeling, to clas-
sify the variables of interest by their orders of integration. This is a task that is by
no means an easy one to accomplish, even with our fairly advanced understanding
of the asymptotic theory, given the low powers of most available tests for unit
roots. I propose a conservative testing strategy to allow for the possibility of
incorrect classification and also deal with the issue of estimating cointegrating
relationships in single equations, dealing with, in particular, the two-step method
proposed by Engle and Granger (1987) and suggesting a simple alternative. This
is linked with the issue of testing for cointegration and unit roots, an area that has
generated considerable interest in the literature. Finally, I consider single-equation
versus systems methods of estimation and show that the choice between these two
methods can be made and understood within the familiar concepts of exogeneity.
The final section concludes. Since frequent reference is made in the text to the
concepts of weak and strong exogeneity, an appendix contains a discussion, based
on Engle, Hendry, and Richard (1983), of these concepts.
DYNAMIC SPECIFICATION AND COINTEGRATION 475
Spurious regressions
While Nelson and Plosser's paper provided one of the early surprising insights in
the literature on unit roots, Yule (1926) had already alerted the profession to the
potential dangers of undertaking stationary inference in an environment with
nonstationary variables. He tenned the phenomenon "nonsense" regressions and
showed how regressing one variable that followed a random walk on another
totally unrelated random walk led to findings of significant correlations between
the two series. Granger and Newbold (1974) returned to the Yule example, and
their fonnulation of the problem fonns the starting point for our analysis. Granger
and Newbold called such regressions "spurious," and this has come to be regarded
as the more commonly accepted terminology.
Granger and Newbold considered the following data-generation process (DGP)
for the data series {X'}~I' {Y/};=I:
Y,=Y,_I + U" u, - llD(O, O'~), (12.1)
Thus X, and Y, are uncorrelated random walks. Before proceeding to the fonnal
details of the example some important features of the data generation process may
be noted. The process generates two variables, each with a unit root. This tenni-
nology can be understood more readily by rewriting the DGP in lag polynomial
operator fonn:
where the processes generating {u,} and {v,} remain unchanged and L is the lag
operator such that Ljx, = X,_j, Ljy, =Y,_j'
The "unit root" in the {x,} and {y,} processes refers to the value of unity for
the coefficients PI and P2' Values of PI and P2 such that IpI! < I, and Ip21 < 1
correspond to "stationary roots." In the terminology of Engle and Granger (1987),
(12.1) and (12.2) are called processes "integrated of order 1", denoted I (1 )-that
is, they need to be differenced once to achieve stationarity. 1
It is interesting to compare the properties of the series {x,} and {y,} when PI
= P2 = 1 with the properties of the series generated by values of Ipll < 1 and Ip21
< 1. In the fonner case both series have unconditional variances that grow with
time (at rate t) while the series have time-invariant finite variances in the latter.
476 MACROECONOMETIUCS
{x,}, also a random walk and therefore also highly persistent, and spurious corre-
lations arise. Further, the use of the nonnal table asymptotically is based on the
assumption that {e,} is a white-noise process under Ho. This is clearly false and
it follows then that the t-statistic is not asymptotically nonnally distributed.
Phillips (1986) also demonstrated an important feature of the Durbin-Watson
statistic (DW) calculated from the residuals of (12.3). When the regression is
spurious, DW ~ 0 in probability. This is a consequence of precisely the property
discussed in the previous paragraph. Under H o, {E,} far from being a white-noise
process is instead highly correlated and this is revealed in a low value of DW.
When the two series are genuinely related, the DW statistic converges to a nonzero
value and the behavior of the DW statistic therefore provides a way of discrimi-
nating between genuine and spurious regressions. We will return to this issue in
a later section.
Inconsistent Regressions
Another example of the dangers involved in using standard distributions for infer-
ence when there are nonstationary variables present was highlighted by Mankiw
and Shapiro (1985, 1986) in their discussion of what have come to be called
"inconsistent" or "unbalanced" regressions.
In this tennino10gy, a regression is said to inconsistent (unbalanced) if the re-
gressand is not of the same order of integration as the regressors or any linear com-
bination of the regressors. The inconsistency (imbalance) refers to the disparity in
the orders of integration of the variables on the two sides of the regression. Thus
the problem occurs if, say, the regressand is an 1(0) variable while the regressors,
individually and in combination, are 1(1). The problem also appears if the re-
gressors are only near-integrated-that is, PI in (12.1 ') is close to, but not equal
to, one in absolute value. Mankiw and Shapiro's analysis, which we describe
below, concentrates on these near-integrated cases, but their results apply equally
well to integrated series.
The difficulty with these inconsistent regressions can be understood in the
context of spurious regressions. An 1(0) variable cannot be related in any mean-
ingful sense to an 1(1) variable (where the 1(1) variable may be taken to be a linear
combination or composite of several 1(1) variables), given their very different
statistical properties and behavior over time. Thus an inconsistent or near-incon-
sistent (if only near-integrated variables appear) regression may be regarded as a
special kind of spurious regression. The use of standard tables will therefore lead
to misleading inferences on the significance of estimates of parameters.
This discussion has considerable economic interest since tests for rational
expectations, in consumption (Flavin, 1981) or the stock market (Fama and French,
478 MACROECONOMET~CS
1989), typically give rise to such regressions. For example, Flavin's (1981) test of
Hall's (1978) random walk hypothesis for consumption takes the form of regress-
ing differenced consumption «(1 - L)c, = ..1c,) on lagged income (Yl-1)' If both
consumption and income are 1(1) variables, this regression is inconsistent. Under
the null hypothesis, that consumption follows a random walk, the coefficient on
the lagged income term should not be significantly different from zero. However,
given the form of the regression, the t-statistic for the coefficient estimate on
lagged income does not have the standard t-distribution and use of the t-table will
lead to spurious findings of significance and hence rejections of the random walk
hypothesis.
Mankiw and Shapiro (1986) consider the following DGP for hypothetical {c,}
and {y,} series:
!i.c, = v"
y, = eY,_1 + e"
E,_I(v,) = E'_I(e,) = 0,
corr(v" e,) = p,
6\p-+
J. Model (12.5) Model (12.6)
1.0 0.9 0.8 0.5 0.0 1.0 0.9 0.8 0.5 0.0
T= 50
0.999 30 24 20 11 7 60 45 36 16 6
0.99 26 20 15 to 7 54 40 33 15 6
0.98 22 17 15 8 7 50 37 30 14 5
0.95 17 12 to 7 6 38 30 25 12 6
0.90 12 9 8 6 6 28 22 19 10 6
0.00 5 6 6 5 5 6 7 7 5 6
T=200
0.999 29 23 20 10 5 61 48 38 18 5
0.99 18 15 13 8 4 41 32 27 13 5
0.98 13 10 9 7 5 29 24 20 11 6
0.95 9 7 7 6 5 17 14 12 7 6
0.90 7 6 6 6 6 to 9 8 6 7
0.00 5 4 4 5 5 5 5 4 5 5
Notes: DGP: (12.4); sample size = T; number of replications = 1.000.
rejections in the upper tail of the density. For values of 8 well within the unit
circle, the size distortions disappear.
Second, the distortions are a decreasing function of T although this holds only
when the series are near-integrated. For 8= 1, there will be no reductions in size
distortions when the sample size increases.
Finally, the distortions are an increasing function of the number of nuisance
parameters (such as constant or trend) estimated. Thus the distortions in model
(12.6) are higher than those in model (12.5). This is again a characteristic feature
of such densities.
In summary, the results in this section again show the dangers of conducting
inference using standard tables when unit roots are present. This leads naturally
to the conclusion that tests of economic hypotheses of interest, such as the excess
sensitivity of consumption to income, crucially depend on a preclassification of
the variables of interest by their orders of integration, as this determines what
critical values should be used for inference. Incorrect preclassifications will lead
to an inappropriate choice of critical values and thus lead to incorrect inferences.
For most of the next section we duck this issue of possible incorrect classifi-
cation and proceed conditionally on a classification of the orders of integration of
480 MACROECONOMETRICS
the variables. We ask the question whether, given that some of the variables have
been classified as 1(1), it is possible under any circumstances to return to using
standard tools of inference and whether such a return has any other practical
benefits such as unbiased or efficient estimation. We then return briefly to the issue
of ameliorating the consequences of incorrect classification by being conservative
in our testing strategy.
Dynamic Regressions
Overview
The fundamental point, which is the first major theme of this survey, emerging
from the discussion above is the striking difference that may arise between the
critical values required to conduct inference in a stationary environment and those
required when unit roots are present, and, as a corollary, the mistakes that can
occur if incorrect critical values are used.
The classification of time series by their integration properties, say, into 1(0)
and I( I), is an area fraught with difficulty. There is a vast literature just on testing
for unit roots, with a wide variety of tests proposed. Each of these tests may have
satisfactory power properties against a given set of alternative hypotheses but may
be powerless against a range of other alternatives.
However, given that the classification has been properly made, series inte-
grated, say, of order one and related to each other seem to offer a special ad-
vantage to the applied econometrician. The asymptotic property that confers this
advantage is called superconsistency, and discussion of this property takes us to
the second important theme of the literature on I(l) processes-namely, the
modelling of long-run economic relationships by means of static regressions.
Engle and Granger (1987) introduced the notion of cointegration to the main-
stream of applied and theoretical econometric research. The idea is simple yet very
powerful and can be understood by looking at a simple DGP (taken from Engle
and Granger):
y, + f3x, = Ut (12.7a)
y, - ax, = e, (l2.7b)
(12.7c)
Earlier in the chapter we noted that the variance of x" if x, is generated by a unit
root process, grows with t. From this, and from the properties of e I' it is possible
to see after some tedious manipulation that the variance of the numerator of (12.8)
is of order T 2 while the denominator has variance of order T 4• Thus to prevent
explosive (or degenerate) behavior of both the numerator and the denominator,
they need to be scaled by T and T 2 , respectively.5 Simplifying, it is then clear that
T(a- a) has a nondegenerate distribution.
The scaling is one important way in which the presence of unit roots alters the
asymptotic theory of the distributions of estimators even in models as simple as
a a
(12.8). tends in probability to a, denoted ~ a at rate T, instead of the usual
T 1/ 2 ,6 the rate of convergence of consistent estimators in stationary asymptotic
theory. This is known as super- or T-consistency, as distinct from T 1I2-consistency.
The integratedness of the series feeds into the distribution of the estimator in
another important respect. The denominator of (12.8) does not tend in probability
to a finite limit but has an asymptotic distribution. Similarly, the distribution of the
numerator of (12.8) is not asymptotically normal. Standard central limit theorems
do not apply because of the nonstationarity of X,.
Finally, consider taking the Cochrane-Orcutt transform of (l2.7b). This yields
(12.9)
Equation (12.8) differs from (12.9) only in the former's omission of an 1(0) term
given by (Y,_I - aX,_I) (1(0) because Y, and X, are cointegrated), which thereby
enters the residual of the regression in (12.8). Nevertheless, a remarkable conse-
quence of T-consistency is that the omission of 1(0) terms in the model does not
affect the consistency of the coefficient estimator on an 1(1) variable. Thus, with
a long data series, a static regression such as (12.8) is enough to obtain a consistent
estimate of the long-run relationship between the variables. However, the finite-
sample and asymptotic distributions of the coefficient estimator is affected by the
presence of unmodeled terms, an observation that is of some importance in our
later discussion. This is also a point taken up by Zivot in his comment, which
follows this chapter.
Engle and Granger's paper emphasized the value of static regressions in an en-
vironment where the processes are integrated of order 1 and recommended modeling
such integrated processes in two stages. In the first stage, the long-run relationship
is estimated via static regressions. At the second state, the dynamics of the model,
all parameterized as 1(0) variables, are estimated. Thus in an 1(1) environment,
static regressions formed, according to this line of argument, an important part of
DYNAMIC SPECIFICATION AND COINTEGRATION 483
good modeling practice. How much of this recommendation still holds true, in the
light of subsequent research, is a matter of some debate. Several interlinked issues
are important in any evaluation of this point. These include, in particular, issues
concerning the distributions of the coefficient estimators, biases in these estimates
and tests for cointegration based on these estimates. We address all these issues
in tum.
Asymptotic Theory
disposable income variables, ~I' .•. , ~P' have been expressed as coefficients on
1(0) variables (given that y, and c, are cointegrated). Because it is possible to
achieve this rewriting, the distributions of the coefficient estimators of ~I' . . . ,
~P' denoted ft h ••• , ftp' are individually and jointly asymptotically normally
distributed. Thus standard normal tables can be used to test for significance of
the individual ftis while F-tables can be used to conduct tests of joint significance
(asymptotically). This therefore represents a considerable simplification of the
process of inference although it is based, importantly, on a proper classification
of the integration and cointegration properties of the variables concerned. This
analysis alone is enough to reinstate the use of dynamic regressions and should
lead to a focus away from static regressions. But there are at least two other
important reasons-the first concerning biases in estimates of the cointegrating
relation and the second linked with the issue of testing for cointegration.
Biases
The discussion in this section is based on the important, but clearly not always
realistic, assumption that single-equation methods are statistically valid (and
efficient) for estimating the parameters of interest. The parameters of interest
include those describing short-run behavior and also those giving the long-run
relationships among the variables. The claim in this section is that, even when
single-equation estimation methods suffice,? both the long-run and the short-run
parameters are better estimated using dynamic methods. Thus the important
notion, that the errors should form a martingale difference sequenceS in a well-
specified model, continues to hold true when dealing with nonstationary series.
Therefore, in finite samples, the biases in the estimates of the long-run parameters,
introduced by not explicitly modeling the dynamic 1(0) terms, which thus enter the
residuals of a static regression, are considerable.9
We illustrate the argument with an example taken from Banerjee, Dolado,
Galbraith, and Hendry (1993) (referred to henceforth as Banerjee et a/., 1993).
While an important charge that can be made against any example is that the DGP
is too special, the results here are representative of a large number of Monte Carlo
results (Hendry and Neale, 1987; Stock, 1987; Phillips and Hansen, 1990).
Suppose the series {YI} and {XI} are generated by the following process:
(12.1la)
X, = X'_I + f 2 ,; (12.1lb)
f ll - NID(O, ai), f 2, - NID(O, aD, COV(fl " f 2s ) = 0, "if t, s;
Yl + Y2 + Y3 = 1.
DYNAMIC SPECIFICATION AND COINTEGRATION 485
Y. = 0.9, Y2 = 0.5
(1/(12 = 3 -0.39 -0.25 -0.15 -o.Q7 -0.04
YI = 0.9, Y2 = 0.5
(1/(12 = I -0.32 -0.22 -0.14 -0.08 -0.04
Y. = 0.5, Y2 = 0.1
(1/(12=3 -0.23 -0.13 -o.Q7 -0.03 -0.02
Y. = 0.5, Y2 = 0.1
(11/(12=1 -0.21 -0.12 -0.06 -0.03 -0.02
Source: Baneljee et al. (1993, table 7.3).
Notes: DGP: (12.1 la) + (12.1 Ib); 5,000 replications. Standard errors of these estimates vary
widely. but the estimated biases are in almost all cases significantly different from zero. The biases do
r
not decline at rate T but they do decline more quickly than n. The simulations used GAUSS. The
mean biases are computed as [(5,OOOr'2,;'looo (a, - 1) l. where i is the index for the replications for each
parameter configuration.
Thus, both {Y,} and {x t} are I( 1) series and are CI( 1, 1), with a long-run multiplier
of 1 linking the two series. to Further, {x t} is strongly exogenous for the regression
parameters.
Consider now estimating the long-run by means of the static regression:
The extra lagged term is included to avoid imposing homogeneity on the relation
between y and x. Although homogeneity would be a valid restriction in this case-
that is, d =0, the extra term allows for the possible ignorance of the investigator
and, as results reported in Banerjee, Galbraith, and Dolado (1990) show, does not
affect the estimate of the short-run adjustment coefficient c. The estimate of the
long-run multiplier is deduced from (12.13) as (1 - d/C).13 For the same configu-
ration of parameters given in Table 12.2, the biases in the estimate of the long-run
parameter derived from (12.13) are all insignificantly different from zero.
It is possible to extend this set of experiments to allow for weakly exogenous
X,. The results carry through in this case. However, as we show in a later section,
if weak exogeneity fails to hold, the usefulness of estimates from dynamic single
equations is reduced substantially and the comparison between static and dynamic
estimates becomes ambiguous. Systems estimation is the main route to consider
in the absence of weak exogeneity and Monte Carlo evidence here (see, for exam-
ple, Gonzalo, 1994) again supports the claim of substantial inferiority, in general,
of estimates derived from static regressions.
Finally, it is important to note that it is possible to derive analytical expressions
that explain the difference in the accuracy of the estimates using alternative
methods. From these analytical expressions, it is possible to derive the parameter
configurations or DGP specifications for which, say, static regressions are likely
to outperform (or perform as well as) dynamic regressions or vice versa (see
Kremers, Ericsson, and Dolado, 1992).
This paper has taken as its main theme the importance of dynamic regressions in
a nonstationary environment. This theme has been emphasized via several sub-
themes discussing the particular benefits of conducting estimation and testing in
a dynamic setting. This section discusses the third sub-theme of this line of argu-
ment, testing for cointegration and discusses first using a specific simple example
and then more generally tests of cointegration in dynamic models.
A large class of tests for cointegration (Phillips and Ouliaris, 1990) takes as its
starting point a static regression such as (12.12) and tests for unit roots in the
estimated residual series U,. A popular test in this category is known as the aug-
mented Dickey-Fuller test and consists of estimating the regression
(
residual series and hence provides prima facie evidence of the existence of
cointegration.
The critical values used for this test are essentially modified Dickey-Fuller
critical values. The unmodified critical values, given in Fuller (1976), apply to
testing for unit roots in raw series. These critical values have to be adjusted for size
when the series to be tested for a unit root is constructed or derived from a re-
gression such as (12.12). Naturally this implies that the critical values are sensitive
to the number of variables in the cointegrating regression and to the existence of
a constant or a trend in the regression. 14 An extensive set of tables is provided by
MacKinnon (1991), while Engle and Granger (1987) and Engle and Yoo (1987)
provide a more limited set.
Return now to the experiment described in (12.11). The parameter configura-
tion given by YI = 1, Y2 = Y3 = 0 generates the case where y, and x, are not
cointegrated. Equation (12.11a) can be written in error correction form as
(12.15)
Thus a test for cointegration can be based on the I-statistic for cin (12.13) because
under the usual null hypothesis of "no cointegration", C = 0 from (12.15). The
distribution of this test statistic is nonstandard because under H o, (Y,_I-
X,_I) is 1(1) while both differenced terms are 1(0) and the regression, in the termi-
nology discussed above, is inconsistent. However, the test is straightforward and
would be useful if it had good power properties.
It is interesting to remark that under the alternative hypothesis of cointegration,
Ic=o is asymptotically normally distributed. This follows from the property that in
this case C is a coefficient on an 1(0) variable and the arguments discussed above
apply.
The 1<=0 test is a simple example of what Boswijk and Franses (1992) call a
Wald test for cointegration and is based on a dynamic regression model with a
constant included. It is instructive to compare the power properties of this test with
an ADF(I)15 test based on the residuals of the static model (12.12). The critical
values for both these tests are derived by simulating the model under the null (that
is, YI = I, Y2 = Y3 =0, S =G 1!G 2 = 1 in (12.11a) and (l2.11b» for 5,000 replica-
tions. These critical values are then used for deriving the test powers, when the
null hypothesis of no cointegration is false, of the tc: o test and the ADF(1) test.
Critical values are reported in Table 12.3 while powers of the tests are given in
Table 12.4. Both tables are taken from Banerjee et al. (1993) and the simulations
used PC-NAIVE (Hendry, Neale, and Ericsson, 1990).16
Several issues may now be noted. First, the most significant divergences in the
powers of the two tests appear in the last three blocks of Table 12.4 and may be
understood in terms of the common-factor restriction imposed by an ADF(1) type
of test. The ADF(I) test involves testing YI = 1 in the regression
488 MACROECONOMET~CS
=
YI 0.9. Y2 =0.5. (1/(12 = I
T= 25 0.14/0.11 0.06/0.05 0.01/0.01
50 0.2 l/O. 15 0.10/0.09 0.0210.02
100 0.49/0.30 0.30/0.19 0.08/0.04
=
Y, 0.9. Y2 =0.5. (1/(12 = 1/3
T= 25 0.13/0.10 0.07/0.05 0.0210.01
50 0.24/0.13 0.1210.07 0.03/0.01
100 0.59/0.24 0.40/0.14 0.13/0.03
YI = 0.5, Y2 =OJ. (1/(12 = 3
T= 25 0.66/0.35 0.45/0.20 0.1610.05
50 0.99/0.84 0.97/0.72 0.78/0.34
100 1.00/1.00 1.00/1.00 1.00/0.97
Y, = 0.5. Y2 = 0.1. (1,/(12 = 1
T= 25 0.79/0.31 0.66/0.18 0.29/0.04
50 1.00/0.80 1.00/0.67 0.94/0.28
100 1.00/1.00 1.00/1.00 1.00/0.96
Y, = 0.5, Y2 = 0.1. (1/(12 = 1/3
T= 25 0.94/0.23 0.87/0.12 0.64/0.03
50 1.00/0.75 1.00/0.60 1.0010.22
100 1.00/1.00 1.00/1.00 1.00/0.94
Table 12.6.
51 l(l) l(l) 5
52 1(1) 1(0) >5
53 1(0) 1(0) 5
54 1(0) 1(1) <5
p-I
where 1t' = 1l(1, -9'). Equation (12.17) is a generalization of the test given in
(12.15). As before, Il = 0 implies that there is no cointegration while if Il ;t 0,
(12.17) can be reparameterized as an autoregressive distributed lag model of order
p with Y, and x, cointegrated. If 9 were known, a test for cointegration could be
based on a t-test of Il = O. In general, this test is not implementable because 9 is
not known but must be estimated. Thus the regression must either be
reparameterized as
L\y, = Co + '~L\x, + Il(Yt-l - t'X,_l) + ~'X'_I
p-I
where t is a vector (of dimension equal to the dimension of XI) whose elements
are all equal to 1 and ; = 1..(1,' - 9'), and the t-test now based on the t-statistic for
~ in (12.18), or, using the result that .A. =0 implies 1t =0, based on a Wald-type
statistic of the form
(12.19)
Systems Estimation
It is natural to ask, given that the discussion above has focused exclusively on
single-equation estimation techniques, how the arguments generalize to estim-
ation in systems. In essence, this is equivalent to asking what happens when the
regressors are not weakly exogenous for the cointegrating parameters. Based on
our discussion so far, it should not be surprising to observe that the long-run
relation is poorly estimated not only in static regressions but that single-equation
dynamic models in general do not perform much better.
There are at least four interrelated issues in single-equation estimation that are
worth highlighting. First, the presence of unit roots introduces nonstandard distri-
butions of the coefficient estimates. Second, the errors may be processes that are
autocorrelated. Third, in a multivariate setting, there can exist several cointegrating
relations and there is no longer a natural ordering of the variables (which depend-
ent and which independent) in a static regression. 17 Finally, the explanatory vari-
ables in the single equation may not be weakly exogenous for the parameters of
interest.
492 MACROECONOMETRICS
(12.21)
DYNAMIC SPECIFICATION AND COINTEGRATION 493
where (U11' U2,)' follow a jointly stationary process. Then, in ECM fonn, the
system can be written as
Conclusion
We have taken as our main theme of this paper the observation that the funda-
mental methods per se of econometric modeling, for specification and estimation,
remain unaltered in the new world of nonstationary econometrics. Some of the
dramatic simplifications that appeared possible, in tenns of, say, modeling the
long-run and short-run separately, or focusing primarily on the long-run properties
of and interrelationships between series, seem eventually to be fraught with dif-
ficulties of tests having low powers and estimators having large biases.
We have also argued that, even for modeling the long run or testing for
cointegration, dynamic models provide the most effective way of obtaining infor-
mation. Where weak exogeneity is violated, single-equation estimation techniques
are, in general, inadequate, and systems-based methods are optimal. Thus the
importance of this literature lies rather in the new awareness we have of the
494 MACROECONOMET~CS
properties of common time series and of the consequent need to take account of
these properties in estimation and to modify, wherever necessary, the tools re-
quired to conduct inference.
The first element of the product on the right side of the equality is known as the
conditional density (or model), while the second element is the marginal density.
In the simple case where X, is bivariate normal, the conditional model leads to a
bivariate regression with y, as the dependent variable and z, the regressor.
Suppose our parameter of interest is given by If! (in the conditional model).
Weak exogeneity of z, for If! requires that (I) If! is a function of ~ alone and (2)
there are no cross-restrictions between A. 1 and ~. The essential element of weak
exogeneity is that ~ contains no information relevant to discovering A. 1• Infer-
ence concerning If! can be made conditional on z, with no loss of information
relative to what could be obtained using the joint density of y, and z,.
Strong exogeneity requires that z, is weakly exogenous for If! and
D(z, I X'_I> A. 2) = D(zr I Z,_I' Yo, A. 2),
DYNAMIC SPECIFICATION AND COINTEGRATION 495
so that Y does not Granger-cause Z. (That is, Y does not enter the process gener-
ating Z. In a simple regression model this implies that the equation generating z,
does not contains any lags of y,.) Strong exogeneity is necessary for forecasting
that proceeds by forecasting future z's and then forecasting y's conditional on the
predicted z's.
Finally, z, is superexogenous for 'If if and only if z, is weakly exogenous for 'If
and dA/dA2 = O. Superexogeneity is necessary where models will be used for
policy analysis, in which the investigator is interested in predicting the derivative
of the dependent variable with respect to a change that can be made in an explana-
tory variable; a change made to the value of the explanatory variable by an external
actor is in effect a change in the parameters of the process governing that vari-
able's evolution.
Acknowledgments
This paper is part of a program of research with which I have been associated for
several years. During this period I have benefited greatly from discussions with
my colleagues Juan Dolado, John Galbraith, and David Hendry. While absolving
them of any responsibility for the views expressed in this paper, it is a pleasure
to acknowledge their help. The financial assistance of the Economic and Social
Research Council (UK) under grant R231184 is also gratefully acknowledged.
Notes
I. This definition can be extended to cover variables integrated of order d > I. Roughly speaking.
a series is said to be integrated of order d if it is stationary after differencing d times but is not stationary
after differencing only d - I times.
2. An issue of some interest is which of model (12.5) or (12.6) is appropriate. The advantage of
using a model such as (12.6) lies. generally. not in its plausibility but rather in making the critical
values appropriate for testing the null H o ; ~ = 0 invariant to the presence or absence of a constant in
=
the process generating cr If this DOP has a constant. say. ACt d + v" the critical values of the t-test
in (12.5). used for testing Ho• are sensitive to the value of d whereas this is not the case in model (12.6).
A related problem concerns what critical values should be used. West (1988) shows that if d te- O.
td'boO is asymptotically normally distributed in (12.5) while it has a nonstandard distribution in (12.6).
In the absence of a priori knowledge of whether or not there is a constant in the DOP. invariance of
the critical values is a useful property and I would argue in favor of using the more general model. and
hence nonstandard critical values. given by (12.6).
3. Source; Mankiw and Shapiro (1986. table 2). The standard errors of each of the entries avo
expressed as a fraction. is given by [(a,)(1 - a,)/N]tl2 where N is the number of replications in the
=
Monte Carlo (N 1.000 for this table).
4. Work on nonlinear cointegration. although mathematically more complex. shares this main idea
of some function of the component series being bounded statistically.
496 MACROECONOMETIUCS
5. By order T' we mean that the variance grows at rate T k • For example, for a random walk process
such as (12.1), it may be shown that variance (YT) =TO'~ and thus k= I in the tenninology given above.
Thus if we define a variable ZT = T- II 2yp variance (ZT) = 0'; which is finite, bounded away from zero
and remains constant with time. Scalings such as these are important in deriving the asymptotic theory
because it ensures that the asymptotic distributions of the scaled expressions are neither nondegenerate
(do not collapse on a single value) nor explosive (have infinitely large variances).
6. That is, it is T(fr. - a), which has a nondegenerate, nonexplosive distribution asymptotically
rather than T,n(fr.- a). The latter is the scaling appropriate for estimators based on stationary processes.
7. It is important to make this qualification in order not to mislead the reader. There are many
circumstances, some described below, especially under failure of weak exogeneity, when single-
equation dynamic models also provide badly biased estimates. However, this does not run counter to
our main assertion here that static regressions are hardly ever desirable and thus two-step methods
based on first estimating the long-run and then modeling the short-run have severe limitations. For
estimation and inference, the two steps are best accomplished together.
8. A martingale difference sequence (MDS) generalizes the concept of an identically and inde-
pendently distributed sequence of random variables, by allowing for some amount of dynamic depend-
ence among the elements of the sequence. An MDS is defined with respect to an information set !JH
of data realized by time 1 - I. A sequence {y" 1 = I, 2, ... } is defined to be a MDS with respect to
{!Jr 1= I, 2, ... } ifE{ly,l} < "" 'V' t and thatE{y,1 !J,_,} =0 'V' I. In words, we require that the expectation
of the absolute value of y, be finite and that the expectation of y, condi/ional on Ihe past be zero. For
the purposes of the analysis that follows, the reader will not lose by thinking of the simpler special case
of serially uncorrelated, mean zero, variables.
9. There is a literature on correcting, nonparametrically, the estimates derived from a static regres-
sion in order to account for the effects of the omitted dynamic terms (see, for example, Phillips and
Hansen, 1990). Provided single-equation estimation is valid, dynamic specification and nonparametric
correction of the estimates from a static regression are asymptotically equivalent procedures. Compari-
sons in finite samples are usually ambiguous and depend on the particular specifications of the DGP.
10. The long-run multiplier is derived by setting y, = y*, x, = x* 'V' 1 and E" to its expected value
of O. Thus, y* = y,y* + (Y2 + YJ)x*, which implies that (1 - }\)y* = (y, + y,)x*. Under the restriction
that y, + )1 + Y, = I, it follows that y* = x* in the long-run.
II. This may be seen by rewriting (12.11 a), using the homogeneity restriction, as y, = x, + y,( Y-
X),_l + (~ - l)l\x, + E".
12. Error-correction models (ECMs) are the focus of an extensive literature, starting wih Sargan
(1964), Hendry and Anderson (1977) and Davidson, Hendry, Sma. and Yeo (1978). They are a way
of capturing adjustments in a dependent variable that may depend not only on the level of some
explanatory variable but also on the extent of the deviation of the explanatory variable from an
equilibrium relationship with the dependent variable. Thus, if the equilibrium relationship is given by
y* = 8x·, the error-correction term is given by (y, - 8x,). In (12.13) 8 = 1. The estimate of the
coefficient on the error-correction term. given above by c. provides an estimate of the short-run
adjustment to equilibrium.
13. If it is necessary to have estimates of the standard error of this multiplier estimate, (12.13) can
be estimated more conveniently in a linearly transformed form that gives the long-run multiplier
directly. This is known as the Bewley transform and the estimates derived from this transform are
numerically equivalent to those obtained from (12.13).
14. See footnote 2.
e
15. ADF(1) refers to an augmented Dickey-Fuller test with = I in (12.14).
16. Critical values for the ',.0 test for the case where (12.13) includes a constant and x, is a vector
(at most of dimension 5) are given by Banerjee, Dolado, and Mestre (1993). These critical values
depend on the dimension of the system (number of regressors). Hansen (1992) provides an ECM test
DYNAMIC SPECIACATION AND COINTEGRATION 497
that is invariant to dimension, but the test imposes a common factor restriction. Where this restriction
is violated, the losses of power involved are important. Banerjee, Dolado, and Mestre (1993) provide
a full discussion of this issue.
17. Techniques that rely on rotating the dependent variable in sequence are in general unsatisfactory.
18. Naturally, this is only a sufficient condition for such a violation. To derive all the necessary
conditions for weak exogeneity would require us to look at the properties of the error process (e'l' eZI)"
19. To emphasize the point made in the earlier footnote, this is clearly only a necessary condition
for weak exogeneity since the parameters of the two equations may be linked in other ways.
20. For details of the first method, see Phillips and Hansen (1990), Phillips and Loretan (1991),
and Phillips (l99\). For the second method, in addition to the original account contained in Johansen
(1988), also see Banerjee and Hendry (1992) and Banerjee et af. (1993).
References
Engle, RF., and B.S. Yoo. (1987). "Forecasting and Testing in Cointegrated Systems."
Journal of Econometrics 35, 143-159.
Fama, E.F., and K.R French. (1989). "Business Conditions and Expected Returns on
Stocks and Bonds." Journal of Financial Economics 25, 23-49.
Flavin, M.A. (1981). "The Adjustment of Consumption to Changing Expectations About
Future Income." Journal of Political Economy 89,974-1009.
Fuller, W.A. (1976). Introduction to Statistical Time Series. New York: Wiley.
Gonzalo,1. (1994). "Five Alternative Methods of Estimating Long-Run Equilibrium Re-
lationships." Journal of Econometrics 60, 203-233.
Granger, C.W.J., and P. Newbold. (1974). "Spurious Regressions in Econometrics." Jour-
nal of Econometrics 2, 111-120.
Hall, R.E. (1978). "Stochastic Implications of the Life-Cycle Permanent Income Hypo-
thesis." Journal of Political Economy 86, 971-987.
Hansen, B.E. (1992). "A Powerful Simple Test for Cointegration Using Cochrane-Orcutt."
Working Paper No. 230, Department of Economics, University of Rochester.
Hendry, D.F. (1992). "On the Interactions of Unit Roots and Exogeneity." Manuscript,
Nuffield College, Oxford.
Hendry, D.F., and GJ. Anderson. (1977). ''Testing Dynamic Specification in Small Simul-
taneous Models: An Application to a Model of Building Society Behaviour in the
United Kingdom." In M.D. Inrilligator (00.), Frontiers of Quantitative Economics (ch.
8c, pp. 361-383). Amsterdam: North-Holland.
Hendry, D.F., and AJ. Neale. (1987). "Monte Carlo Experimentation Using PC-NAIVE."
In T. Fomby and G. Rhodes (eds.), A Century of Economics (Vol. 6) (pp. 91-125).
Greenwich, CT: JAI Press.
Hendry, D.F., AJ. Neale, and N.R. Ericsson. (1990). PC-NAIVE: An Interactive Program
for Monte Carlo Experimentation in Econometrics. Oxford: Institute of Economics and
Statistics, Oxford University.
Johansen, S. (1988). 'Statistical Analysis of Cointegration Vectors." Journal of Economic
Dynamics and Control 12, 231-254.
Kremers, J.J.M., N.R Ericsson, and J. Dolado. (1992). "The Power of Co-integration
Tests." Oxford Bulletin of Economics and Statistics 54, 325-348.
MacKinnon, 1.G. (1991). "Critical Values for Co-integration Tests." In R.F. Engle and
C.WJ. Granger (eds.), Long-Run Economic Relationships (pp. 267-276). Oxford: Oxford
University Press.
Mankiw, N.G., and M.D. Shapiro. (1985). "Trends, Random Walks and Tests of the Per-
manent Income Hypothesis." Journal of Monetary Economics 16, 165-174.
Mankiw, N.G., and M.D. Shapiro. (1986). "Do We Reject Too Often? Small Sample
Properties of Tests of Rational Expectations Models." Economics Letters 20, 139-145.
Nelson, C.R., and c.I. Plosser. (1982). "Trends and Random Walks in Macro-Economic
Time Series: Some Evidence and Implications." Journal of Monetary Economics 10,
139-162.
Phillips, P.C.B. (1986). "Understanding Spurious Regressions in Econometrics." Journal
of Econometrics 33, 311-340.
Phillips, P.C.B. (1991). "Optimal Inference in Co-integrated Systems." Econometrica 59,
282-306.
DYNAMIC SPECIFICATION AND COINTEGRATION 499
Phillips, P.c.B., and B.E. Hansen. (1990). "Statistical Inference in Instrumental Variables
Regression with 1(1) Processes." Review of Economic Studies 57, 99-125.
Phillips, P.C.B., and M. Loretan. (1991). "Estimating Long-Run Economic Equilibria."
Review of Economic Studies 58, 407-436.
Phillips, P.C.B., S. Ouliaris. (1990). "Asymptotic Properties of Residual Based Tests for
Cointegration." Econometrica 58, 165-193.
Phillips, P.c.B., and W. Ploberger. (1991). "Time Series Modelling with a Bayesian Frame
of Reference: I. Concepts and Illustrations." Manuscript, Yale University.
Sargan, J.D. (1964). "Wages and Prices in the United Kingdom: A Study in Econometric
Methodology." In P.E. Hart, G. Mills, and I.K. Whitaker (eds.), Econometric Analysis
for National Economic Planning. London: Butterworth. Reprinted in D.F. Hendry and
K.F. Wallis (eds.), Econometrics and Quantitative Economics. Oxford: Basil Blackwell,
1984.
Sims, C.A., I.H. Stock, and M.W. Watson. (1990). "Inference in Linear Time Series with
Some Unit Roots." Econometrica 58, 113-144.
Stock, I.H. (1987). "Asymptotic Properties of Least-Squares Estimators of Co-integrating
Vectors." Econometrica 55, 1035-1056.
Stock, I.H. (1992). "Deciding Between 1(0) and 1(1)." Technical Working Paper No. 121,
National Bureau of Economic Research.
Stock, J.H., and M.W. Watson. (1988). "Variable Trends in Economic Time Series." Jour-
nal of Economic Perspectives 2, 147-174.
Stock, J.H., and K.D. West. (1988). "Integrated Regressors and Tests of the Permanent
Income Hypothesis." Journal of Monetary Economics 21, 85-96.
West. K.D. (1988). "Asymptotic Normality When Regressors Have a Unit Root." Econo-
metrica 56, 1397-1418.
Yule, G.U. (1926). "Why Do We Sometimes Get Nonsense Correlations Between Time
Series? A Study in Sampling and the Nature of Time Series." Journal of the Royal
Statistical Society 89, 1-64.
Commentary on Chapter 12
Eric Zivot
Introduction
501
502 MACROECONOMETRlCS
u;
Let Y; =(Yl,' Y2,) be a two-dimensional I(l) process and =(u l " U2,) be a stationary
and ergodic 1(0) process. We assume that Y, is cointegrated with a triangular
structural representation
(la)
(Ib)
IT,) T I
T- 1/2 Lu, => B(r), T-3/2L~' => JB(r)dr,
I I 0
J J
T I T I
where [Tr] denotes the integer part of Tr and => denotes weak convergence (the
function space equivalent of convergence in distribution).
Our interest is on the efficient estimation of f3 in the system (la) and (lb). Let
P denote the ordinary least squares (OLS) estimator of f3 using (Ia) as a regression
model. Then it can be shown that
504 MACROECONOMET~CS
P
r
~ U1I,(,)'d, UII, (,jdB,W ",}
This result demonstrates that converges to the true value f3 at rate T instead of
(2)
the usual rate T I12 and that t3 scaled by T converges to a nonnormal random
variable.
The expression for the limiting distribution in (2) is not very convenient for
analysis since the Brownian motions Bir) and BI(r) are correlated (n is non-
diagonal). Phillips (1989) shows, however, that we may orthogonalize Bir) and
BI(r) through the relation
where '1'= C021C02~ and B I2 (r) is a scalar Brownian motion, independent of B2(r),
with variance CO Il _2 = COli - CO~IC02~. Using (3) we may decompose the limit dis-
To gain some insight into the above expression, consider separately the three
terms in (4). The first integral term is a mean zero and symmetric "mixture nor-
mal" random variable. That is,
fi
distribution of is polluted by asymmetries, nuisance parameters (co2 1> (0 22 ) and
a bias term (~21)' The presence of these effects makes the asymptotic distribution
fi
of nonnormal and invalidates the use of conventional test statistics (e.g., t and
F tests) and tables for inference about f3.
In the special diagonal case
L = ["'II(L)
VIC ) 0
0] L
Vl22(L) ,
= [0'11
0
0]
0'22
so that Y2t is strictly exogenous, it is easily shown that C0 21 =.121 =0 and the limit
distribution in (4) simplifies to the mixture-normal random variable
(5)
where W2(r) and Wdr) are independent standard Brownian motions and CO II .2 =
COli' Since the distribution in (5) is mixture-normal, standard test statistics can be
used for inference about f3.4 Moreover, in this special case, the limit distribution
(5) is the same as that obtained from applying full systems maximum likelihood
estimation (MLE) on (1a) and (1b) that explicitly incorporates the short-run dy-
namics embodied in ,¥(L).5 We have the interesting result that when Y2t is strictly
exogenous, OLS on the static regression model (1a), which ignores the short-run
dynamics in U 1t, produces an estimate of f3that is asymptotically equivalent to full
systems MLE.
As was mentioned above, dynamic specification is not the only method available
to augment the static cointegrating regression to improve inference. We now
illustrate how the technique of FM-OLS works to correct the static regression
equation. FM-OLS corrects the static regression for the long-run endogeneity be-
tween Y21 and Ylt and for the contribution of the serial correlations in the error terms
U 1, and U2t to the long-run covariance of U t so that when OLS is applied to the
corrected static regression the second and third terms of the nonnormal limiting
distribution in (4) are eliminated and the (mixture) normal limiting distribution (5)
obtains. Simple modifications of conventional test statistics can then be made so
that inference about f3 can proceed using standard tables.
The modification of the static regression proceeds in two steps. First, the long-
run endogeneity between Y2t and Ylt is purged through the transformation
(6)
506 MACROECONOMET~CS
where 'If = (021 COiL which creates a diagonal long-run covariance matrix for the
random vector (ur" U2,)'. Second, any remaining asymptotic bias due to residual
correlation between ur, and U2, is captured by constructing
(8)
(9)
which is the asymptotic distribution of the full systems MLE (5). It is important
to note that this result holds regardless of whether Y2t is weakly exogenous for /3.
The usefulness of the convergence result (9) is that inference about /3 can
proceed with standard test statistics with a simple modification to take account of
the long-run variance (0112' For example, a modified t-ratio has the form
iJ = L
T-\
w
(. )
L t(j), ,& = LW L
T-I (.)
tV), tv) = r- I LU,U;-j,
T
the regression whereas a dynamic specification tackles the problem using paramet-
ric corrections. FM-OLS has an asymptotic advantage over dynamic specification
if Y2' is not weakly exogenous for {3, since FM-OLS explicitly corrects for long-
run endogeneities. However, dynamic specification may have a finite sample
advantage over FM-OLS in situations where weak exogeneity holds due to its
parametric structure. This is due to the fact that a dynamic specification estimates
parametrically the short-run dynamics as well as the long-run coefficient {3 while
the FM-OLS procedure only corrects for the impact of the short-run dynamics on
the structure of the long-run covariance matrix. The simulation evidence reported
in Phillips and Hansen (1990), Hansen and Phillips (1990), Phillips and Loretan
(1991), and Stock and Watson (1993) generally support these claims.
Consider a special case of the triangular cointegrated model (Ia) and (lb) where
we assume a stable first order VAR structure for u,: C(L)u, =E" E, - iid N(O, I),
C(L) =[ - CL and C has elements cj •j U, j = I, 2). In this case, we may write u,
=(I - CLf' E, = 'Y(L)E,. The long-run covariance matrix of u, is given by Q =
'Y(I)I'Y(I)' = ([ - CrII([ - C'r l . More explicitly,
Here, u l , =C'IUlt-l + Ell' u2J = E21' so that only U I , is serially correlated but UI, and
U2' are contemporaneously correlated. To derive the single equation dynamic speci-
fication for this model we must reduce the system (la) and (lb) by conditioning
(la) on Y2, and [t-!. First we rewrite (Ia), using (Ib), to produce the system
508 MACROECONOMETRICS
[
VI ] -1.I.d.
• . N ((0)' .(<7'~1 <7'~2)) (11)
£2/ 0 <7'12 <7'22
with <7'11 * = <7'11 + {32<7'22 + 2{3<7'12 and <7'12* = <7'12 + {3<7'22' Then, using properties of
the bivariate normal distribution, it is easy to show that
where J'I = C11' Yl ={3 + 1/>, Yl =-({3c 11 + 1/». Notice that (13) is consistent with the
long-run equilibrium YII = {3Y21 since (1'2 + Yl)/(l - YJ) = {3.
Using {3 = (Yl + Y3)/(1 - YJ), we may also write the dynamic specification as
(14)
Equation (14) is the static regression equation (1a) augmented with the lagged
cointegrating term (YI - {3Y2),-1 and .1Y21' Both of these terms are 1(0) (i.e., station-
ary) and serve as parametric corrections for the long-run endogeneity of Y2' and the
serial correlation in "It. However, due to the nonlinear restrictions on {3 in (14),
OLS cannot be used to estimate {3.
We may also express the dynamic specification as the error correction model
(ECM):
(15)
From (15) we see that if Yl = I, there is no error correction term and the model is
expressed in first differences so that Ylt and Y2' are 1(1) and not cointegrated.
Banerjee discusses how OLS on (15) can be used to both test for cointegration and
to estimate {3.
Equations (13), (14), or (15) together with (1b) produces an orthogonal system
with error structure
111) _ 1.I.d.N,
( £21
.. ((0) (<7'112
0 0
0
<7'2
J) . (16)
COMMENTARY ON CHAPTER 12 509
From inspection of the conditional model (12), the only parameter common to
both distributions is (122' Clearly, the marginal model for Y21 contains no additional
information about 13 beyond what is available in the conditional model (12) and
so Y2t is weakly exogenous for 13.
With C I2 = C21 = C22 = 0, the OLS estimate of 13 from the static equation (la)
has the asymptotic distribution given in (4) with '1/= if>(I - cllf l and .1 21 = 0'21(1
- cllf', so that this estimate suffers from asymptotic asymmetry and bias. Notice
that these effects are caused by the nonzero value of (112' the contemporaneous
correlation between U lt and U2,' On the other hand, since the dynamic specification
(14) is a valid conditional model the (nonlinear) least squares estimate of 13 from
(14) is asymptotically equivalent to the full systems MLE. In addition, since the
short-run dynamics are explicitly estimated conventional test statistics can be used
to test hypotheses about 13 without modification.
If the data are in logs, we can interpret the coefficient on .1Y2t in (15), r2,as
the short-run elasticity with respect to Y't and f3 as the long-run elasticity. Notice
r2
that since = if> + f3, Y2 =f3 only if 0'12 =O. In this case (la) reduces to the common
factor model
where Cit is independent of f1t. 7 As noted above, in this case Y2t is strictly exogenous
and OLS on the static regression model (la) is asymptotically equivalent to full
systems MLE.
This special case of the triangular representation helps shed some light on the
biases in {3 from the static regressions reported in Table 12.2 from Banerjee. In this
simulation experiment, Banerjee used the ADL model (13) as the OGP and in all
configurations of the parameters reported in Table 12.2 the common factor restric-
tion is violated. Therefore, the OLS estimate of 13 from the static regression (I a)
when the DGP is (13) will suffer from both asymptotic and finite sample biases.
Further, since r2 = if> + f3, the magnitude of the bias should increase with the
magnitude of if> (that is, with the difference between the short-run and long-run
elasticities) and the simulation results confirm this observation.
The FM-OLS estimator is easily constructed in this case. Since '1/ = tro'2i, the
first step modification, which makes U ll long-run uncorrelated with U2" is simply
510 MACROECONOMETRICS
This regression is similar to the dynamic specification (14). The dynamic speci-
fication, however, includes a lagged error correction term that eliminates the first-
order serial correlation in u ll , whereas the FM regression ignores the short-run
effects of the serial correlation. In addition, in the dynamic specification 1], is
uncorrelated with UZ1 by construction but in the FM regression U;, is only long-run
uncorrelated with U2/' Thus there might be some finite sample biases in the FM-
OLS estimator, relative to the dynamic specification estimator, of f3 due to short-
run correlation and endogeneity effects.
where 1], = YI, - E[YIt I YZI' I,-tl and is uncorrelated with ~, by construction. In
(18), the effects of the serial correlation in UZ 1 is removed by the addition of the
regressors LlYzI and .1YZI_1 to (la).
COMMENTARY ON CHAPTER 12 511
where Yz =P+ 4', Y3 =-(I + C22)4', 'r.l = C224'· Notice that (20) is consistent with
the long-run equilibrium YIr = PY2r since Yz + Y3 + 'r.l = p. The ECM fonn is given
by
(21)
Again, the short-run elasticity Yz is equal to the long-run elasticity Ponly if 0'12 =
O. In this case, however, Y3 = Y4 = 0 and the model reverts to (1a) with Y21 strictly
exogenous.
From (19), the marginal model for Y2r is
and we see immediately from inspection of the conditional distribution (18) that
Y2r is weakly exogenous for p. Thus the OLS estimate of p in (18) will be asymp-
totically equivalent to full systems MLE and standard test statistics can be used
without modification for inference.
OLS on the static regression equation (1a) is asymptotically biased since'll =
4'(1 - en) and ~21 = 0'12' As in Case I, this bias is caused by the contemporaneous
correlation between "Ir and "21' If 0'12 = 0, then OLS on (1a) is asymptotically
equivalent to full-systems MLE, but standard test statistics need to be adjusted for
the long-run variance of "2' before they can be used for inference about p.
To determine the FM-OLS estimator we first fonn
The FM-OLS estimator is then given by substituting (22) and (23) into (8).
It is interesting to compare the FM-OLS procedure to the dynamic specification
(18). In (18), the addition of ~Y21 and ~Y2r-1 to the static regression (la) are
sufficient to remove the asymptotic bias tenns. The modification in (22) involves
only ~Y21 and so the bias correction tenn in (23) is needed to pick up the effects
of ~Y2'-I'
512 MACROECONOMETIUCS
where TJ, and E21 are uncorrelated by construction. Interestingly, the conditional
model (23) has the same form as the conditional model (14) for Case 1, where
there is no feedback from U It- 1 to U Z, ' This result suggests that (23) does not take
into consideration the feedback between the errors in (la) and (lb). However, the
marginal model for YZI' (24), is different from that in Case 1 since (24) retains the
feedback from UI,_I to UZ I through the lagged error correction term.
The levels specification of (23) is the ADL model
Clearly, from (27), {3 directly affects the marginal distribution of YZI through the
error correction term unless CZI = O. This means an estimate of {3 from conditional
model (24) does not use all of the information in the system regarding {3 and will
COMMENTARY ON CHAPTER 12 513
be less efficient than an estimate utilizing all infonnation in the conditional and
marginal models. Therefore, YZt is not weakly exogenous for 13.
In this example, we have the interesting result that CZ I ::I: 0 implies that the error-
correction tenn enters the marginal model for h" which results in the failure of
weak exogeneity of YZt for 13. More generally, Urbain (1993) shows that the
absence of the error correction tenn in the marginal model of the conditioning
variables turns out to be a necessary and sufficient condition for valid inference
on the long-run parameters using a single equation conditional model derived
from a cointegrated system.
The bias tenns in the asymptotic distribution of {3 are given by
These effects are caused by two factors: the contemporaneous correlation between
"It and "Zt captured by 0'12; and the feedback from "It-I to Uzt caused by CZI' In
particular, the conditioning process used to derive (23) removes the correlation
between "11 and "Zt so that O'ZI = 0 (see (16)) but it does not remove the feedback
from "1'_1 to "Z,' In other words, the conditioning process fails to diagonalize the
long-run covariance matrix Q. Therefore the asymptotic distribution of the
(nonlinear) least squares estimate of 13 from (23) is given by (4) with If! =
CZIWll(CZ/Wll + O'zzt and a ZI = Jr. As long as CZI::I: 0 (ht is not weakly exogenous
J
the biases in the first stage cointegrating regression have a smaller effect on the
power of residual-based tests for cointegration than the proper specification of the
second step ECM regression.
Acknowledgments
Notes
References
Hendry, D.F. (1994). "On the Interactions of Unit Roots and Exogeneity." European Uni-
versity Institute Working Paper ECO No. 94/4, Florence.
Johansen, S. (1988). "Statistical Analysis of Cointegration Vectors." Journal of Economic
Dynamics and Control 12, 231-254.
Newey, W.K., and K.D. West. (1987). "A Simple, Positive Semi-Definite Heteroskedasticity
and Autocorrelation Consistent Covariance Matrix." Econometrica 55, 703-708.
Park, 1. (1992). "Canonical Cointegrating Regressions." Econometrica 60, 119-143.
Park, J.Y., and M. Ogaki. (1991). "Inference in Cointegrated Models Using VAR
Prewhitening to Estimate Shortrun Dynamics." Rochester Center for Economic Re-
search Working Paper No. 281, University of Rochester.
Phillips, P.C.B. (1988). "Reflections on Econometric Methodology." The Economic Record
64, 344-359.
Phillips, P.C.B. (1989). "Partially Identified Econometric Models." Econometric Theory 5,
181-240.
Phillips, P.C.B. (1991). "Optimal Inference in Cointegrated Systems." Econometrica 59,
282-306.
Phillips, P.C.B., and RE. Hansen. (1990). "Statistical Inference in Instrumental Variables
Regression with 1(1) Processes." Review of Economic Studies 57, 99-125.
Phillips, P.C.B., and M. Loretan. (1991). "Estimating Long-run Economic Equilibria."
Review of Economic Studies 58, 407-436.
Stock, J.H., and M.W. Watson. (1993). "A Simple Estimator of Cointegrating Vectors in
Higher Order Systems." Econometrica 61, 783-820.
Urbain, J.P. (1993). Exogeneity in Error Correction Models. Lecture Notes in Economics
and Mathematical Systems 398. Springer-Verlag.
Zivot, E. (1994). "Single Equation Conditional Error Correction Model Based Tests for
Cointegration." Institute for Economic Research Discussion Paper 94-12, Department
of Economics, University of Washington.
13 NONLINEAR MODELS OF
ECONOMIC FLUCTUATIONS
Simon M. Potter
Introduction
The modern study of economic fluctuations rests on stylized facts produced by the
use of linear time-series models. I There are certainly important controversies con-
cerning the best type of linear models and the interpretation of stylized facts from
these linear models as illustrated by some of the other chapters of this book. How-
ever, as a class linear models have a number of serious shortcomings for the study
of economic fluctuations. 2 They impose a symmetry of the effect of shocks across
stages of the business cycle and signs and magnitude of shocks. All fluctuations
produced by linear models are generated by exogenous forces and typically these
exogenous forces appear to be large. Nonlinear models have the advantage of be-
ing able to capture asymmetries over the business cycle and in allowing the inter-
nal dynamics of the economy itself to be one of the ultimate sources of fluctuations.
Economists are often taught that "there is no such thing as a free lunch," thus
it must be the case that nonlinear models have some disadvantages. Indeed, the
very disadvantages of nonlinear models are the advantages of linear models. Non-
linear models are difficult to specify and estimate, and perhaps more important,
once estimated it is hard to directly pin down their important dynamics. Linear
models can all be derived from one common specification and have widely available
517
518 MACROECONOMETIUCS
estimation programs, and their global and local dynamics can be obtained directly
by reference to the dynamics contained in their impulse response functions and
variance decompositions. The first purpose of this chapter is to clearly exposit the
tradeoffs between linear and nonlinear models.
If one takes a purely statistical view of the tradeoff between linear and nonlinear
models, then the choice between them should be based on the result of a statistical
comparison. An alternative way of stating this is that since linear models are the
dominant tool used in economics, then "evidence for nonlinearity" is required be-
fore nonlinear models become a commonly used tool. The second purpose of this
chapter is to discuss at a nontechnical level some of the issues involved in testing
for nonlinearities. Readers requiring a more technical discussion are directed to
Brock and Potter (1993) and Granger and Terasvirta (1993).
The chapter then proceeds to discuss a general class of parametric nonlinear time-
series models-endogenous delay threshold autoregression-that are proving useful
in describing asymmetries in economic variables over the business cycle. In the
last few years a number of univariate nonlinear models have been proposed and
estimated for U.S. output. The most well-known is due to Hamilton (1989), but
the concentration will be on the recent models due to Beaudry and Koop (1993)
and Pesaran and Potter (1994). The latter two is a member of the endogenous delay
threshold autoregression (EDTAR) class and works off the intuition of floor and
ceiling effects in the economy producing nonlinearities.
The chapter concludes by providing a guide to the use of impulse response
functions and variance decompositions for nonlinear models. Techniques from
Potter (1996) are described and applied to the floor and ceiling model. Thus, the
third and final purpose of this chapter is to give the reader some idea of existing
and potential future results using nonlinear models.
Background
of criticism started in the classic articles of Yule (1927) and Slutzky (1927) of
earlier econometric work that had modeled business cycles by assuming that the
economy was driven by a periodic process with the same time-series properties
as the economy itself. Jevons's (1884) sunspot theory of the business cycle is the
standard example. In criticizing this work all three (Yule, Slutzky, and Frisch) had
made use of the properties of linear difference equations driven by random distur-
bances. They found that such statistical models were very capable of producing
simulated time series that mimicked the behavior of the actual business cycle.
After World War II the linear time-series techniques they promoted came to
dominate the study of economic time series. The dominance was produced by the
relative sophistication of these new techniques in conjuction with the new feasi-
bility of computation. Further, with the Keynesian revolution in macroeconomics
it was easy to interpret the impulses as "animal spirits" and the propagation mech-
anism as the aCclerator-multiplier mechanism.
The Wold Representation. Five years after Frisch's Cassell paper, Wold (1938)
showed an interesting theoretical property of all time series that are covariance
stationary. It is possible to decompose these series into two components: one
component that was perfectly linearly predictable from the past of the time series;
a second component that is a weighted sum of random variables that cannot be
linearly predicted from the past history of the time series. This second component
is often called the Wold representation by economists. Thus, if {Y,} is a time series
with no perfectly predictable components, its Wold representation has the follow-
ing form:
Y, = LY';U'-i,
i=O
where E[ V,] = 0, E[ V;] = G., E[ VPI-i] =0 if i :# 0, Y'o = 1 and L:.o Y'; < 00.
Some intuition for the Wold representation can be gained from the follow-
ing informal argument. Define the linear predictor of YI+N at time t (using p
observations):
Y, = iPYt- 1 + V"
then
and
N1
Y, = Y, - iPYt- 1 + iPY,-1 - iP2Y,_2 + ... + iP - Y,-N-I - iPNY,_N + iPNYt-N'
which simplifies to
Yt = V, + iPV,_1 + ... + iPN-1V,_N_1 + iPNY'_N'
Under the assumption of covariance stationarity and the lack of any purely
deterministic components in Yt the use of an abitrary finite p in the definition of
the linear predictor will approximate (in the mean square convergence sense) the
optimal predictor over the infinite past (see Brockwell and Davis, 1987, for more
formal statements). Thus, as p ~ 00 one could write Y, as a sum of linear prediction
errors based on forecasting using information in the infinite past,
where Lt-;[Y,) = L/-i[Yt I YH , Yt- i- I, .. ,]. Alternatively, one could follow the in-
tuition of the AR( 1) example above, and attempt to find an equivalent expression
in the one step ahead forecast errors, Lt-i[Y/-i] - L,-i-I[Y,-i)' The Wold represen-
tation is just such a expression, where the sequence of one-step forecast errors are
multiplied by coefficients given by a simple Fourier transformation to provide
optimal linear predictions of Yt. The power of the Wold representation is that no
form was assumed for the time series other than the properties of covariance
stationarity and the lack of deterministic linear components, The latter feature can
be incorporated by adding a deterministic sequence that is covariance stationary
to the Wold representation.
NONLINEAR MODELS 521
Y, = L'P;U /-i ,
i=O
where E[U,] =0, E[U,U;] =L,., E[U,U;_d =Ok if i::t 0, 'l'o =I k and I,;'o II 'f'; Ib
< 00.
is far more difficult to solve. There are two main difficulties. First, there is no
general theory for developing the nonlinear predictor. Second, the nonlinear pre-
dictor is indexed by N above to indicate that the nonlinear predictor for N = 2
cannot necessarily be obtained from the nonlinear predictor for N = 1. For the
moment I write the solution to this problem for emphasis as NL[·], but later I will
use the conditional expectation notation, E[.]. Only in the case of a linear model
will the linear predictor agree with the conditional expectation.
For the first problem, one possibility is to restrict attention to a particular class
of nonlinear models:
simplifications produced in the linear case by the chain rule of forecasting. Sup-
pose we have available L,[ Y,+n), n = p, ... ,N - 1. Then, using the linear predictor
we have
L'[Y'+N) = a + ~)iL'[Yt+N-;]'
;=1
With a nonlinear predictor, the chain rule of forecasting still applies but it does
not lead to any simplifications. In the case that Y, has a Markov property-that is.
the conditional distribution of Y, depends on a finite number of the past values of
the time series-it is possible to generate the n step ahead forecasts from the one-
step ahead conditional distributions. For example, suppose
Y, = g(Y,_I) + V,.
Then
NL[Y,+I I Y, = y,) = J(g(y,) + v,+I)/vdv,+lo
NL[Y,+2 I Y, = y,) = J[J(g(y) + vt+2)/v dvt+2 }t:vtv,dY,
where Iv is the density of the innovation V, and the conditional density hly, is equal
to Iv with all values translated by the value of g(y,):
n=1
In the case that Y, has a Markov property of order p, then the nonlinear predic-
tors in (13.3) will contain all the relevant information about the dynamics of
NONLINEAR MODELS 523
nonlinear models in a similar manner to the way the linear predictors in the Wold
representation contains all the relevant information about linear models. If Y, does
not have a Markov property, then it is not obvious how to construct unique meas-
ures of the dynamics. The above considerations lead to the condition that the
analysis of nonlinear models be restricted to those satisfying a Markov property.
The dynamics of estimated linear models are mainly evaluated by the use of
impulse response functions and variance decompositions.
The Univariate Case. There are four possible conceptual experiments one can
go through to generate the impulse response function of a linear model:
• Compare two realizations of the same covariance stationary time series from
time t onwards. One realization, {l~}, is hit by an "impulse" u, at time t and
no impulses thereafter. The other realization, {Y,}, is hit by a zero impulse
at time t and no impulses thereafter. Using the Wold representation one has
Y,iN = LV'iUI+N-i,
i=N
YI+N = LV'iUI+N-i,
i=N+1
• The difference between two linear predictions of the series with one infor-
mation set perturbed from the other by the shock U at time t. Again using the
Wold representation:
• The difference between two linear predictions of the series one from time t
the other from time t - 1:
choosing between measures of the dynamics of the system there is the further
difficulty that now history and shocks matter. This leads to a reporting problem.
It is possible to find histories and shocks that can generate arbitrary responses.
Below I discuss ways of reporting representative infonnation on dynamic re-
sponses in stochastic nonlinear models.
• Use prior infonnation to identify some underlying shocks that are independ-
ent of each other and set all but one shock to zero,
• Orthogonalize the shocks using a Cholesky factorization and set all but one
shock to zero, and
• Integrate out over all the shocks but one as described below. This method
will work for nonlinear models as well.
Without going into the details of any of the procedures, it is the case that the
result of all of them is a K x 1 pertubation vector Vk for each shock and the (K x
1) impulse response function of the kth shock is then given by
Once the method of combining shocks is decided the impulse response function
is uniquely determined in the linear case. In a vector time-series model it is also
of interest to ask which shock is causing the most variability in each variable. The
technique used to answer this question is a variance decomposition. A variance
decomposition decomposes the forecast error variance of each variable into con-
tributions from each of the K shocks in the model. Using the Wold representation
again one can write the forecast error variance as
n
+ VAR(L'+'[Y'+N] - L'[Yt+N])'
Hence, the variance decomposition of the linear model gives information on the
relative importance of the impulse response functions defined by particular com-
binations of the underlying shocks. As shown in the fifth section such a relation
is not true for the nonlinear case.
An initiating impulse, when it comes into play, operates upon a certain complex
of industrial and monetary conditions. Given the impulse. these will determine
the nature ofthe effect that it produces. and are, in this sense, causes ofindustrial
fluctuations. The impulse is the dropping of a match: the consequences are de-
termined by the nature of the material with which it comes in contact
In order to illustrate the differences between nonlinear and linear cases a Volterra
series expansion is introduced. This is a special form of nonlinear moving average
that allows one to give some intuition for the additional effects present in nonlinear
NONLINEAR MODELS 527
models. In order to keep the discussion tractable only the univariate case is
examined:
with V, iid symmetrically distributed and the coefficients 'If;j' 'If;jk also have a
symmetry property.
I will concentrate on the third order expansion and assume that the innovation
sequence {V,} is directly observed. Thus, in this special case the nonlinear predic-
tor is given by
+ LLLV'ijkEt-n[~-i~-j~-k]'
i=1 j=l k=l
which simplifies to
n-l n-I 00
impulse to have the same effect for different initial conditions. In order to illustrate
these characteristics, consider the nlirffor the Volterra series expansion above:
2 3
nlirf,.(v, VI' VI_I, VI_2'" .) = V'lin + v '11nn + v '11nM + 2v L'IInjV,-j+N
j=n+l
+ 3v 2 L'IInnjV,-j+n + 3v L L'IInjkV,-j+nV,-k+n
j=n+1 j=n+1 k=n+l
n-I
+ 3V0'2 L'IInjj'
j=1
• The magnitude of the shock for the same history will produce asymmetries
because of the presence of v2 , y3.
• The response to V will not be the same as -v because of the v2 terms.
• The same shock will have differing effects depending on the sequence of v"
VI_I, ....
Assume that V, « O. It is possible for certain histories that the direct negative
effect is outweighed by the squared term and the terms not involving v,.
• Asset prices sometimes fluctuate a great deal after the announcement of news
that had been predicted.
• Stabilization policy can only have a second-order effect at best on the welfare
of the representative consumer (see Lucas, 1987).
In the first case it is often assumed that the price of a stock depends on expec-
tation of future dividends and interest rates. Assume that the interest rate is con-
stant. The change in the value of a stock is given by
where P, is the ex-dividend price of the stock, D, is the dividend, and 0 < p < 1
is the discount factor.
Suppose that dividends follow an integrated process. If
- 1 - tpn+1
p, - E,_I[P']= Lpn v,.
n=1 l-l/J
Hence, in the case that v, is small, the change in stock price is small and zero if
v, is zero. If alternatively the change in the dividend is described by the cubic
Volterra model, then simple calculations lead to the updating function:
n-I
In the special case that v, = 0 and V,_s = 0, S > 0, we would have for the
unanticipated movement in the stock price
n
n
p, - E,-I[P,l = (12 L p L'I'i;'
n=1 i=1
530 MACROECONOMETIUCS
The asset price has an unanticipated movement even though the current change in
dividend was perfectly predicted. In the nonlinear case, there is "news" in this
perfect prediction that a linear predictor cannot extract.
Lucas (1987) considers the value of stabilization policy by asking how much
initial consumption a representative agent would be prepared to give up to remove
fluctuations in consumption around "trend" for the rest of time. He finds that the
amount is second order compared to the consumer's willingness to pay for an
increase in the growth rate with initial consumption. For linear models, Lucas's
experiment can be interpreted as reducing the variance of the time-series innova-
tion in the consumption stochastic process to zero. Suppose instead that the change
in the logarithm of consumption is actually generated by the cubic Volterra series
model. First note that the average growth rate of consumption will be approx-
imately given by the mean of this process:
E[~lnCt) = (}"2
LlJf;;.
;=\
Next assume that the growth rate is positive. Reducing the variance of the inno-
vation V, would decrease growth and therefore reduce welfare of the representa-
tive agent. Depending on the relative importance of (}"2 versus L %1 V';; for the value
of the mean, this could have a large effect on welfare even for small changes in
the variance.
If we return to the two prediction problems discussed in the previous section, then
the choice between linear models and nonlinear models at first appears obvious.
Choose the nonlinear model if the minimized value of the quadratic loss function
is smaller than that of the linear predictor. A major difficulty is that the linear
functions are a potential choice under the minimization of NL(·). Thus, the loss of
the linear predictor is an upper bound on the loss for the unrestricted predictor. A
similar problem occurs in deciding the order of an autoregression. By increasing
the lag length one can always reduce the squared loss.
There are three main solutions to the question of order selection for linear time
series models:
• Check the residuals of the estimated model for evidence of further linear
structure. For example, one might calculate a Box-Lung statistic for the
residuals.
• Penalize the loss function for the number of parameters that must be estim-
ated relative to the number of observations.
NONLINEAR MODELS 531
The analogs in the case of evaluating linear models versus nonlinear models are
as follows:
• The residuals of linear models are checked for evidence of nonlinear depend-
ence. There are two main ways of doing this:
Use a nonparametric approach such as the BDS test (see Brock, Hsieh,
and LeBaron, 1991).
Use a test function that looks for specific departures from the assumption
of linearity. For example, one common approach is to see if the residuals
are correlated with polynomial functions of the regressors (see Tsay, 1986).
Many economists believe that tests of the significance of the nonlinear tenns
are crucial in deciding whether nonlinear models are worth using. Since econo-
mists have a great deal of sunk capital in linear models, there is much to support
this position despite the possible advantages of nonlinear models in capturing
business cycles dynamics. Note that a statistical test does not give any infonnation
on whether the dynamics of a nonlinear model are economically interesting.
The testing solution has a less direct connection with the problem of
overparameterization in linear time-series modeling. In order to understand the
difficulties consider the likelihood ratio test in the case where for the linear and
nonlinear models a Gaussian innovation is assumed and the linear model is nested
within the nonlinear model. Then the likelihood ratio test statistic for a sample of
size Tis
T(ln( O"L) - In( O"NL) ).
If the linear model is obtained from the nonlinear model by restricting all
parameters that are not shared, then the likelihood ratio statistic will have a
532 MACROECONOMETRICS
Now in order to deal with the dependence across values Hansen's method uses
the empirical distribution of the time series itself, as it is used in the formation of
the test statistic, to appropriately link the realizations of the computer simulated
chi-squared random variables. This leaves one unresolved question: what test
statistic or combination of test statistics are the best to use? In recent work An-
drews and Ploberger (1994) show that in addition to the commonly used maxi-
mum statistic, various weighted averages have good statistical properties.
Nonlinear time-series models of U.S. output share the common theme of regime
switching to model possible nonlinearities. Further, the regimes are taken to rep-
resent different phases of the business cycle providing a connection with the
notion of business-cycle asymmetry. The standard form of regime-switching model
was introduced by Tong and Lim (1980).
Let {X,: t = -00, ... , -1, 0, 1, ... ,} be a time series and let 1, be an indicator
random variable taking values in the set {1, 2, ... , K}. Then the canonical thresh-
old autoregression is defined by
X, = a{J,) + 1/>{J,I(L)X,_1 + (i{J,IV" (13.5)
where V, is an iid sequence of standardized random variables with zero means and
unit variances; and for 1, = j, ali) is a constant, I/>lil (L) is a finite-order polynomial
in the lag operator L.
The main characteristic of the TAR class is its use of piecewise linear dynamic
models over regimes classified by the value of an index variable 1,. In Hamilton
(1989) the regime switching is exogenously generated by an unobserved Markov
chain. In the threshold autoregression models of Potter (1995) and Tiao and Tsay
(1994) the regime switching is endogenously generated by a fixed lag or lags of
output growth. In the model of Beaudry and Koop (1993) an additional variable
enters the autoregression that is only nonzero when output is below its previous
maximum. Thus, Beaudry and Koop's model contains two regimes with endog-
enous switching. We start by working with Beaudry and Koop's model because
of its simplicity compared to the others and its direct economic interpretation.
Beaudry and Koop construct a variable they call the current depth of recession
variable, CDR,:
CDR, =X, - max(X" X,_I, ... , Xo),
where X, is the logarithm of output and Xo is the level of output in 1947Ql.
534 MACROECONOMETIDCS
The first difference of the logarithm of output (approximately the growth rate)
is then modeled as the Beaudry and Koop model:
Y, = a + tP/L)Yt-I + 9CDRH + V"
where Y, = X, - Yt-\ and tPiL) = tP1L + ... tPpU.
Beaudry and Koop find a statistically significant negative coefficient on the
current depth of recession variable. Since they do not estimate any nuisance para-
meters under the alternative hypothesis, a simple t-test is valid in this case. More
important, the estimated () is negative, indicating that the previous maximum of
output acts as a ratchet effect. The ideas of floors or ratchets in output growth was
very common in the 1950s but has fallen out of favor more recently (see Gordon
and Klein, 1965, for a review of the most important articles). To see exactly why
the negative coefficient represents a ratchet effect we consider the following sce-
nario: current output growth is negative, thus next period the current depth of
recession variable has a positive effect on output growth. Suppose that this upward
pressure on the growth rate is outweighed by a further negative output shock. Now
the current depth of recession variable has a larger magnitude and represents more
upward pressure on the growth rate. This upward pressure on the growth rate will
remain in force until output exceeds the previous maximum.
The dynamics found by Beaudry and Koop are similar to those of the threshold
autoregression model of Potter (1995) and Tiao and Tsay (1994). The dynamics
of output suggest that the economy tends to have an in-built stabilizer that prevents
recessions from lasting as long as implied by the forecasts of linear models.
Further, the growth in the recovery phase is higher the deeper was the recession,
the more negative the CDR, variable. A stylized fact that is supported by other
nonparametric methods. One feature that the model of Beaudry and Koop does not
contain, found in previous work particularly Hamilton (1989), is a sharp drop into
recession. The construction of Beadry and Koop's model does not allow for such
effects. In the more general floor and ceiling model of Pesaran and Potter (1994)
this effect is allowed as well as pairing the floor of Beaudry and Koop with a
ceiling. In order to do this we return to the canonical formulation of the threshold
autoregression model above.
{
a K +lPK(L)X,-1 +O'KV, (13.6)
Since the index variable is constructed from the location of observable lags of
the time series, the specification is relatively easy to estimate, test, and evaluate
as exemplified in Tong (1990). For example, in Potter (1995) the second lag of
quarterly output growth, Y,-2, is used as the delay variable with the threshold at
zero growth to define a two-regime model. Although the model fits the data well
and significantly better than a linear model, it is difficult to give a convincing
economic rationale why only the second lag should be so crucial. Further, the
specification of abrupt changes in behavior based on the value of output growth
at a two-quarter lag is difficult to reconcile with nonlinear theoretical models
where one would expect the most recent output growth rate, Y,-l' to have nonlinear
effects on Y, as well. Tiao and Tsay (1994) go some way toward dealing with this
criticism by introducing two regimes conditional on Y,-2 < 0: (1) one regime where
output growth continues to worsen, Y,_I < Y,-2 and (2) a second regime where
output growth improves in the most recent quarter, Y,-t> Y,-2'
One can continue in the manner of Tiao and Tsay and introduce further re-
gimes. For example, it is possible to further split the two contractionary regimes
above depending on whether Y,-3 < 0 or not. However, this is not a parsimonious
approach, and in applying it one could soon run out of degrees of freedom when
analyzing most macroeconomic time series. One possible method of avoiding this
difficulty is to utilize parameter restrictions across the various regimes. The EDTAR
class provides a natural method of generating such restrictions by an endogenous
generation of the index variable defining the regimes and a recursive construction
of additional variables to be added to the autoregression.
Potter (1996) contains a description of the general class. For expositional pur-
poses we concentrate on a somewhat simpler form here which is extended below
for the case of the floor and ceiling model of Pesaran and Potter (1994). We also
introduce a generalization of the indicator function, the squashing function. Squash-
ing functions take on values in [0, 1] as does the indicator function, but they allow
for a smooth transition between the value of 0 to 1.
First, a collection of M index variables, called endogenous delay variables {Iml'
m;; 1, ... ,M} or {Nml , m ;; 1, ... ,M}, is generated by a feedback relationship
among past values of the time series. The form of feedback relationship has some
similarities to recurrent neural networks (see White, 1993, for an overview of
neural networks). Note that in what follows, X, will be the generic time series from
the start of this section:
536 MACROECONOMETRICS
1m< E
= ~{[!,(XH Am>}- for m = I•.... M. (13.7)
(13.8)
where 1(A) is the indicator function taking the value of unity if the event A occurs
and zero otherwise, Am are nonoverlapping intervals on the line and F(·) is the
logistic function:
1 + exp(-13m(Xt - s - rm »
Notice that as 13m ~ 00 that the logistic function is equal to 0 if X,-S < rm,
1/2 if X,_S = r m' and 1 if Xt _ s > r m.
The collection of endogenous delay indicator variables can then be used to
recursively define new variables. In the indicator function case,
(13.1 l)
(13.12)
Potter (1996) discusses conditions required to ensure that the maximum pos-
sible lag length in the recursion is finite in the case of the indicator function and
the sum itself converges in the squashing function case (see below for some brief
details). Intuitively the conditions relate to the stationarity of the series being
modeled. If the time series is stationary, then it should not remain in one location
for a long time. Thus, the consecutive time periods spent in any of the regions
defined by the thresholds will be finite. The squashing function version of the
NONLINEAR MODELS 537
model clearly gives a more flexible lag structure, but this comes at the expense of
significantly more parameters to be estimated.
Given M endogenous delay variables {llr' 121 , ••• , IMI } and the recursive vari-
ables Zml they generate when applied to the time series {X,}, we can define the
EDTAR(p, M) model by
=
where lor = l( 2.~;1 I ml 0) and (ML) is a pth order polynomial in the lag operator.
Alternatively, one can use the recursive variables Wmt to define the smooth
endogenous delay threshold autoregression, SEDTAR(p, M), model by
M
X, = a + tPiL)X,_1 + L6 W
m;1
m ml- 1 + O'V,.
One advantage that the EDTAR model has over the SEDTAR model is that
heteroskedasticity in the error term can be modeled using the endogenous delay
variables. It is possible to use the endogenous delay variables from the SEDTAR
to model conditional heteroscedasticity but it requires more parameters.
The EDTAR model can be reexpressed to show a more direct relationship to
the TAR using (13.11):
(13.13)
The above extended description of the EDTAR model shows that it has the
ability to parsimoniously capture a large number of possible regimes. The cost is
538 MACROECONOMETIUCS
the restricted way that lags enter the different regimes. The SEDTAR model
allows the lags to enter in a less restricted manner.
The additional flexibility of the SEDTAR model can be illustrated by using
(13.12) to give
X, = a + ,piL)Xt-l
and
Thus, vector time-series versions of the EDTAR and SEDTAR models are
given by
where 4l/L) is a pth order matrix polynomial in the lag operator, 8 m , U m are K
x K matrices, a is a K x 1 vector and Zml = (Zlm" ... , ZKm,)" The vector SEDTAR
model is
NONLINEAR MODELS 539
W mJ = (W 1m , • •••• WKm ,)' and HIi' scales the identity variance matrix of Vr •
Estimation and Testing. Both of these models are very parsimonious compared
to other possible extensions of nonlinear time series models to the vector time
series case. Further. once the Zkmt' Wkm , variables are known in any of the models.
they can be estimated by linear least squares. This is especially important in the
vector time-series case where the total number of parameters is large. For the
EDTAR models the main difficulty in estimation is the need to evaluate the model
at all possible threshold values to form estimates. The lack of smoothness of the
indicator function form of the nonlinearity prevents the use of gradient descent
type algorithms to estimate the thresholds. The SEDTAR can be estimated
by traditional gradient descent methods or the neural network technique of
backpropagation.
In order to test for the significance of the nonlinear terms one must deal with
the nuisance parameter problem. In the EDTAR case the nuisance parameters are
the thresholds defining the sets Am. and the Hansen simulation method can be
applied to the grid used for the estimation. In the SEDTAR case there are more
nuisance parameters. and the Hansen approach requires very large computer re-
sources to apply. One possibility is to consider the EDTAR model for testing and
then check whether the SEDTAR version of the model has superior properties. An
alternative testing approach for the EDTAR model discussed in Pesaran and Potter
(1994) is to use a likelihood ratio test of a linear heteroscedastic model against the
full nonlinear model. When the heteroscedasticity is described by the endogenous
delay variables this test has a conventional chi-squared distribution.
OPXM]
F., •
where Ip is the indentity matrix, 0 is a matrix of zeros with the given dimensions.
and the M x M matrix F, is given by
:
540 MACROECONOMETRICS
o
1(121 > 0)
o ],
o I(1Mt > 0)
The Markov representation can be used to give an expression for the time series
N periods ahead in terms of the innovations between t + 1 and t + N:
Notice that if the innovations are zero after time t, the "stability" of the time
series is determined by {n~=IAl+nB}As shown in Potter (1996) this expression
goes to zero under certain conditions on the first row of B. Using results from
NONLINEAR MODELS 541
Markov chain theory it is possible to show that this condition is sufficient to ensure
stationarity of the time series.
Consider the following reformulation of the CDR, variable of Beaudry and Koop:
(13.18)
Although not included in the EDTAR class described above, the variable F,
constructed above is contained within the more general formulation in Potter
(1996).
One difficulty noted above was that Beaudry and Koop's model failed to cap-
ture the initial strong downward movement in recessions. By amending the recur-
sion defining the CDR, variable one can capture this effect. This is done by including
the floor threshold value in the recursion defining CDR, for the first time that the
floor regime is activated:
The regime where neither the floor or ceiling variables are activated is given
the name of the corridor. The residual variances are also allowed to change be-
tween regimes (as in EDTAR model above).
Thus, we have the following nonlinear floor and ceiling model:
• At time t output has just entered the corridor regime from either the floor or
ceiling regimes.
• Assume that for all s > t that Vs O.=
• Assume that Y1+1 is in the corridor regime.
Then we will have the following expression for future values of output growth:
Thus, under this scenario the long-run growth rate inside the corridor regime
will be given by the entering growth rate and approximately one third of the
entering change in the growth rate. This simplifies to two thirds of Y, and one third
of Y,_I'
These values are themselves mainly determined by the behavior of the over-
heating and current depth of recession variables. Under the conditions of this
experiment the long-run growth rate could be anywhere in the interval [-.876,
0.539].
The validity of this experiment crucially depends on the assumption that all
future shocks are set to zero. As noted in the second section this can be very
misleading in describing the dynamics of nonlinear models, and this is particularly
true in the current case. In the case of the floor and ceiling model the standard
deviation of shocks in the corridor regime is estimated to be large compared to the
width of the estimated corridor regime. One would, therefore, expect movements
within the corridor regime in the short run as well as a high probability of exit from
544 MACROECONONUnlUCS
Table 13.1.
the regime in the medium term. Pesaran and Potter contain an extended discussion
of the dynamics of the floor and ceiling model using some of the techniques dis-
cussed in the next section.
This sections reviews some of my own work, Potter (1996) and Koop, Pesaran,
and Potter (1995) (KPP hereafter) on the extension of the usual linear tools of
impulse response functions and variance decompositions to nonlinear models. For
impulse response functions one can proceed in the same manner as the linear case
except one replaces the linear predictor with the conditional expectation operator
as discussed in the second section. Variance decompositions require a different
approach. A summary of the issues developed in that section is in table 13.1 taken
from KPP.
The generalized impulse response function (GI) is designed to solve the problems
categorized in Table 13.1. It represents a choice of the updating definition of the
impulse response function given in the second section. Let 00'_1 be the set contain-
ing information used to forecast Y,. Under a Markov assumption a sufficient set
of information would be the realizations of Y, from t = t - I, t - 2, ... , t - p,
or for the floor and ceiling model the set of realizations would be (X'_I' X,_2, X'_3'
cdr,_I, oh,_d. The convention that lowercase values represent realizations and
uppercases random variables is maintained with oor-! being used to denote a par-
ticular realization of Q,_I' the set of all possible realizations. If a lowercase letter
is superscripted by 0, then it means the actual observed realization; thus, oo~_1 is
the actual observed time series, v~would be the observed innovations-that is, the
NONLINEAR MODELS 545
residuals of the estimated model. The realization of the random variable at time
t + n, Y'+n' will depend on W,_I and {v" ... , v,+.}. Finally, to emphasize the
difference between a conditional expectation evaluated at a point in the sample
space (that is, a particular realization) and a conditional expectation as a random
variable we use EIY,+. II .Q,_I] for the latter.
First consider the GI for the case of an arbitrary current shock, v" and history,
W,_I:
(13.25)
All the measures defined in (13.23) to (13.26) produce the same information from
the generalized impulse response functions when applied to linear univariate
models. For the linear model
(13.27)
which is independent of nt-l. The dependence of G/ on V" the size of the current
shock, is proportional and can be readily dealt with by appropriate scaling of the
GI function. Once the GI function is scaled by V" it is equivalent to the outcomes
of the various definitions of the linear impulse response function given in the
second section.
In the vector case, recall the infinite moving average form:
Y, = L'I'n V,_
n:O
n•
and, once again does not depend on the history, £1/_ 1, However, since V/ is now
a vector, its effect on the generalized impulse response function cannot be elim-
inated by an "appropriate" scaling of the variables. Hence, the impulse response
analysis is subject to the composition problem. One solution to the problem would
be to view the GI as a function of the random variable, V" and derive its probabil-
ity density function in terms of the density function of V,. For example, in the case
where V, - N(O, Iv), then G/(n, V" .a,_I) - N(O, 'P" Iv'l';), and the impulse response
function is fully characterised by the variances, 'P"Iv'l';, n = 0, 1,2, .... These
variances measure the effect of shocking the system at time t, by a random vector
V" and then averaging the resultant densities across all such systemwide shocks.
This solution deals with the composition problem by ignoring it through focussing
on the effect of system-wide shocks.
The GI also lends itself naturally to a solution to the compositional problem
without the use of a priori theory in both the case of linear and nonlinear models.
We can define the generalized impulse response function to be conditional not on
all the shocks at time t but on just one of them. That is, we can consider fixing one
of the shocks, ViI' from the vector of all shocks, V" and then integrating out the
effects of the other shocks at time t given its value, Vi':
where E[Y,+n II Vi' = ViI' nt-.l means that one is taking the expectations conditional
on each of the realizations in .a'_1 for fixed values of the ith shock at time t while
averaging out all other contemporaneous and future shocks.
In order to illustrate this use of the generalized impulse response function,
consider the Gaussian case from KPP. Assume that the vector of random shocks,
V, is jointly normaIly distributed with zero means and the covariance matrix Iv.
Under this assumption
NONLINEAR MODELS 547
n
Gly(n, Vi" ,ur_l) = ('I'n11i)(
~ Vi' )
~.
Scaling the GI by v;/.J(i;, we obtain the effect of a "unit" shock to the ith
disturbance term on Yt+n-namely,
'l'n11i ei
~
where ei is a selection vector with its ith element equal to unity and zeros else-
where. This impulse response function appropriately takes account of the histori-
cal patterns of correlations between the different shocks as characterized by Iv.
Also, unlike the orthogonalized impulse responses obtained using a Cholesky
decomposition of Iv, the GI responses
'l'n11i ei
~
are unique and are not affected by reordering of the variables in Y,.
Although the measures discussed above have a wide range of uses, in many cases
the nonlinear model under analysis suggests a particular set of conditioning events.
For example, many nonlinear models used in economics imply that impulse re-
sponses in recessionary times are different to those in expansionary times. Attempts
to average out over recessionary and expansionary regimes could produce mis-
leading information. For example, suppose that negative shocks are attentuated in
recessionary times but magnified in expansionary times. And positive shocks are
attentuated in expansionary times but magnified in recessionary times. Then
averaging over histories to compare the effect of positive and negative shocks will
tend to weaken the asymmetries.
For example, in the univariate model of PP above there are three regimes that
output can be in: floor (recession), ceiling (fast growth) and corridor (normal
times). One can then define generalized impulse response functions conditional
on being in one of these regimes at t-l for a particular set of shocks. Figure 13.1
548 MACROECONOMETRICS
N
0
I
0..
0
,..
'-' ciI
u...
0
-.J
W
>
<D
0
1\
W
-.J
I
\
'-'
0
-.J (X)
\
\
0
Z I
w
'-'
Z
- -
0
~
:r: 'I
~V
u
N
l v,= 0.32 r
I 0 4 8 12 16 20 24
HORIZON
Figure 13.1 a
0..
o
'-'
,..
o
I
/ --- --
.............
u...
o
d -
N
/
>
W
I
/
/
-.J
'-' 0
o N
-.J I /
Z
/
w
~ (X)
'"V /
~ N
:r: I
U
<D
1-- v,= -291 I
'"I 0 4 8 12 16 20 24
HORIZON
Figure 13.1 b
CD
a
I
0..
o
<.:) a
\
u... -
o I
--J
W
>
.........---
-
W
--J ....
\
o<.:) -
I / r--
/
--J
1/
t5 <Xl
\-
Z -
« I
I
U
N
1- v,= - OB6 1
N
I 0 4 8 12 16 20 24
HORIZON
Figure 13.1 c
a
0..
~ a
( \
u...
o / \
..J
a / \\
----- -
W
> cr>
w
--J
a
/
/
<.:)
S a
<Xl
~
Z
w I
a
<.:)
Z
« " Ia
G '\/ a
I \It 0691
a
CD
aD 4 8 12 16 20 24
HORIZON
Figure 13.1 d
550 MACROECONOMETItlCS
o -
Q
o.
-
V """--
'-' Ul
l.L 0
o
.....J 0
w
> 0
W
.....J Ul
'-' 0
o I
.....J 0
Z -
I
w
\
r
'-' Ul
Z -
«
- -
I \
-- - -
I
() 0 \
1_-
- 1- -
N pos,t,ve shocks
Ul
N
I
"- - negative shocks
I 0 4 8 12 16 20 24
HORIZON
Figure 13.2a
7
Q N
o
'-' ~
l.L
o
.....J
W
>
W 0
.....J
o _
'-'
.....J I
1\
Z
\
W N
'-' I \
--
Z
«
-- - - - -
I
() '"
I
\
"-
"- ,/
,/
- - --
I- -
-
po,,',ve shocks
negatIve shocks
I
..,.
I
o 4 8 12 16 20 24
HORIZON
Figure 13.2b
ex:>
o
f\
Q.
o <D ~
l.L
o
<.:> 0
..,. \
'\
0
-'
W '"
i::. 0
-' \
\ .- - - -- - -- - - - - - --
<.:>
'3 '"
o
w
Z I
/ '\
<.:>
Z I
\ -
o
<D
~
« 0
I
~
'--
I
1--
- ...
I pos,t,ve shocks }
I- - negotlve shoCks
o. !/ I
I 0 4 8 12 16 20 24
HORIZON
Figure 13.2c
Q.
o
<.:>
l.L
o
<f)
0
1\/ -
o
-' 0
';; 0
w
-' <l)
<.:> 0
o I
-' 0
Z
w
<.:> <l) \
Z
« \
I
- - - - -- - - -
U 0 \
~
'"I 1- -
- - po"t,ve shocks
"
negatIve shocks
<l) ...... .....
'"I 0 4 8 12 16 20 24
HORIZON
Figure 13.2d
NONLINEAR MODELS 553
Variance Decompositions
In the case of linear models there is a relationship between the variance decom-
position and impulse response functions when the Cholesky orthogonalization
approach is used. Often the intuition is given as measuring the size of the effect
554 MACROECONOMETIUCS
o
r-~;;;;;::::----;;;;:=;;;;:::-::---------------,
lX)
cO
()
~
C\Q
J "
lL<J
C
"~
.....
:)"t
.C\"
"~ ()
.....
Vl
"~C\j
IS
of shock i under the assumption that all other shocks are zero from time t onward.
For nonlinear models such a treatment of the future doesn't make sense. The
nonlinearity implies that adding up the individual tenns produced by such an
experiment does not give a realization of the original series.
where we return to our standard convention that E[V, V:J =I K• Next consider a
particular realization of the shocks from time t + I to time t + N, (VtN' ...• Vtl)'
with the special property that
(13.29)
where E'[YI+NJ =E[YI+N I V" lO,_I]. Now use (VtN, ...• vt.) as a point around which
to approximate the random variable Y,+N in a first-order Taylor Series expansion
with a remainder term:
"
Y,+N = Y(V~N' ... , vtl; Y,) + L",,(N, n)(V,+v - v~,II)
"=1
--_
{)YI.,+N ... _--
{)YI.,+N
dvl.l+j {)v K,l+j
{)Y2.I+N
{)VK,,+j ,
and are evaluated at {v11' ...• vtN}' andR(V'+N' ... , V'+I; Y) is the Kx I remain-
der vector from the first-order Taylor series expansion.
The '¥,(N, n) matrices and the vectors vt. are fixed at time t and are not
correlated with the future realizations of the time series. Using this fact and using
the conditional expectation operator, E,[·J on both sides of (13.30), it can be shown
that
N
E,[R(t, N)] = L 'P,(N, n)v,~" ,
"=1
• The heteroskedasticity in the error process. This is captured by 'I',(N, N), the
differential with respect to v,+N evaluated along the (v*(t, n)} sequence. For
the vector EDTAR model from the fourth section above this can be written
as
Notice how the previous period's regime matters both in terms of the deriva-
tive of Yt+N with respect to Yf+N-1 and in terms of the initial propagation of
the shock at time time t + N - 1 as represented by the '1',( N - I, N - I) matrix.
As N increases relative to the point in time at which the derivative is taken,
the familar recursion for generating the moving average representation of a
linear VAR is obtained. The differences are the presence of the additional
{em} matrices for the various regimes that could be activated along the {v*(t,
n)} sequence and the importance of the particular regime in place when the
shock, v*(t, n) occurs. Note that the direct effect '¥,(N, n) does not have to
die out as (N - n) ~ 00 even if the underlying series is stationary. Since the
{v*(t, n) } sequence might imply that the time series becomes stuck in one
of the regimes as N becomes large. For example, this would happen in the
floor and ceiling model of output if the long run-mean of the growth rate of
output is outside the corridor regime. The remainder in the expansion would
cancel out this effect for a stationary time series.
NONLINEAR MODELS 557
For a linear model the third term is zero, and the first and second terms are
usually assumed to be constant across the business cycle. That is, in the linear case
the shocks are propagated by the same moving average matrices for each possible
time period t.
=FE(t - I, N + 1) - FE(t, N)
N
+ R (t - I, N + 1) - R (t, N).
Since the generalized impulse response function does not depend on the values
of the innovations after time t it is possible to set these future values to unit vectors
under the understanding that the remainder term depends on this particular reali-
zation (signified by the use of a lower case, r):
Returning to the case of endogenous fluctuations from the second section we can
see that even if V, is a zero vector the generalized impulse response function might
not be a zero vector because of the difference in the remainder terms and through
differences in 'I't-1(N + I, n + I), 'P,(N, n).
N
L~(N, n)Ik'P,(N, n)' + E,[R(t, N)R(t, N)' - E,[R(t, N)]Et[R(t, N)],]
j~n
The effect of shock k on the forecast error variance of the ith variable is given
by
Intuitively if the remainder term accounts for a large fraction of the forecast
variance for a particular history then this is a sign of a large nonlinearity in the
system.
One could define a collection of variance decompositions by looking at all
possible histories. Indeed because the individual contributions are all between [0,
1] the analysis is a little easier. For example, one would say that shock k for
variable i is the most important if it first-order stochastically dominates the con-
tribution of all other variables or more weakly if its average effect over all histories
is the largest. Alternatively, if the contribution of the remainder term first-order
NONLINEAR MODELS 559
stochastically dominates all the other shocks then this would represent a case
where nonlinearity was the main cause of fluctuations.
Acknowledgments
This paper has benefited greatly from discussions with Woon Gyu Choi, Gary
Koop, and Hashem Pesaran on related issues. Financial support from the NSF
under grant SES 9211726 and the Center for Computable Economics at UCLA is
gratefully acknowledged.
Notes
I. Throughout the chapter the following notational conventions will be used. Uppercase lellers
will represent random variables; lowercase lellers will be realizations of these random variables.
Boldface will indicate matrices or vectors. V" V, will be reserved to represent the random (independent
and identically distributed) innovation to a time-series model. In the case of the scalar V, the innova-
tions have mean zero and unit variance. In the case of the random vector V ~ each component of the
vector has mean zero and unit variance. The individual components of V, are mutually independent of
each other unless specifically noted to the contrary.
2. Note that by linear models we mean a property of the model itself rather than of the estimation
technique for the model. In particular ARMA models are linear in the sense used in this paper but
require nonlinear estimation techniques.
References
Acemoglu, D., and A. Scott. (1993). "Asymmetries in the Cyclical Behavior of U.K. Labor
Markets." Mimeo, MIT.
Andrews, D., and W. Ploberger. (1994). "Optimal Tests When a Nuisance Parameter Is
Present Only Under the Alternative." Econometrica 62, 1383-1414.
Beaudry, P., and G. Koop. (1993). "Do Recessions Permanantly Affect Output?" Journal
of Monetary Economics 31, 149-163.
Brock, W., D. Hsieh, and B. leBaron. (1991). Nonlinear Dynamics, Chaos and Instability.
Cambridge, MA: MIT Press.
Brock, W., and S. Potter. (1993). "Nonlinear Time Series and Macroeconometrics." In
Maddala, Rao, and Vinod (eds.), Handbook of Statistics Xl. Amsterdam: Elesevier.
Brockwell, P., and R. Davis. (1987). Time Series: Theory and Methods. New York: Springer.
Christiano, L. (1990). "Solving a Particular Growth Model by Linear Quadratic Approx-
imation." Manuscript.
Frisch, R. (1933). "Propagation Problems and Impulse Problems in Dynamic Economics."
In Economic Essays in Honor of Gustav Cassell. London: Allen and Unwin.
560 MACROECONOMETIUCS
Gallant, R., P. Rossi, and G. Tauchen. (1993). "Nonlinear Dynamic Structures." Econo-
metrica 61,871-908.
Gordon, R., and L. Klein. (1965). American Economic Association Readings in Business
Cycles. Homewood, IL: Irwin.
Granger, C., and T. Terasvirta. (1993). Modeling Nonlinear Dynamic Relationships. Oxford:
Oxford University Press.
Hamilton, 1. (1989). "A New Approach to the Economic Analysis of Nonstationary Time
Series and the Business Cycle." Econometrica 57, 357-384.
Hansen, B. (1993). "Inference When a Nuisance Parameter Is Not Identified Under the Null
Hypothesis." Mimeo, Department of Economics, University of Rochester.
levons, W. (1884). Investigations in Currency and Finance. London: MacMillan.
Koop, G., H. Pesaran, and S. Potter. (1995). "Impulse Response Analysis in Nonlinear
Multivariate Models." Journal of Econometrics, forthcoming.
Lucas, R. (1987). Models of Business Cycles. New York: Blackwell.
Pesaran, M., and S. Potter. (1994). "A Floor and Ceiling Model of U.S. Output." Working
Paper, Department of Economics, UCLA.
Potter, S. (1990). Nonlinear Time Series and Economic Fluctuations. Ph.D thesis, Univer-
sity of Wisconsin.
Potter, S. (1995). "A Nonlinear Approach to U.S. GNP." Journal ofApplied Econometrics.
Potter, S. (1996). Nonlinear Time Series and Macroeconomic Flactuations. London: World
Scientific.
Slutzky, E. (1927). "The Summation of Random Causes as the Source of Cyclical Pro-
cesses." In The Problems of Economic Conditions. Moscow: Conjecture Institute.
Tiao, G., and R. Tsay. (1994). "Some Advances in Nonlinear and Adaptive Modeling in
Time Series Analysis." Journal of Forecasting 13, 109-131.
Tong, H. (1983). Threshold Models in Non-linear Time Series Analysis. New York: Springer-
Verlag.
Tong, H. (1990). Non-linear Time Series: A Dynamical Systems Approach. Oxford: Oxford
University Press.
Tong, H., and K. Lim. (1980). "1breshold Autoregression, Limit Cycles and Cyclical
Data." Journal of the Royal Statistical Society Series B, 42, 245-292.
Tsay, R. (1986). "Nonlinearity Tests for Time Series." Biometrika 73, 461-466.
White, H. (1993). Artificial Neural Networks. Oxford: Blackwell.
Wold, H. (1938). A Study in the Analysis of Stationary Time Series. Uppsale, Sweden:
Almqvist and Wiksell.
Yule, G. (1927). "On a Method of Investigating Periodicities in Disturbed Series." Trans-
actions of Royal Society, London, Series A 226, 267-298.
Commentary on Chapter 13
Daniel E. Sichel
Introduction
561
562 MACROECONOMETIUCS
Before jumping in, a few preliminaries. In this comment, I focus almost exclu-
sively on the empirical side of this literature. Moreover, most of this research has
examined univariate models of broad macroeconomic indicators, such as gross
domestic product and unemployment. I will do the same.
Establishing Benchmarks
Certain important dynamic (and maybe structural) features of the economy can be
captured only by a fully parameterized nonlinear model. Hamilton's (1989) re-
gime-switching model and Beaudry and Koop's (1993) nonlinear transfer function
model are important examples, as discussed more fully below.
Improving Forecasts
for the sake of improving forecasts but also because forecast performance on data
not used to estimate the model is one of the most rigorous diagnostics to which
a nonlinear model can be subjected.
Methodological Advances
Econometric tools and techniques for nonlinear modeling only come to life when
applied to data. Although some applications may not generate important new facts
about economic fluctuations, they may nevertheless help build a tool box that will
prove useful later.
As Simon Potter points out, the key benefit of nonlinear models is the richness of
possible dynamics, while the principle cost is the additional complexity. Different
assessments of these costs and benefits will lead to different judgments about the
importance of particular research. Nonlinear practitioners may see large benefits
in richer dynamics, but general macroeconomists may see high costs from need-
less complexity. To reduce these costs, nonlinear practitioners may well need to
focus less on technique and more on what the nonlinearities teach us about the
economy.
To see this, consider three papers generating nonlinear results that were
published in top journals and have had a large impact on macroeconomics: Neftl;i
(1984), Hamilton (1989), and Beaudry and Koop (1993).
Neftl;i's paper examined the fundamental question of whether economic pro-
cesses differ in expansions and contractions. Specifically, Neftl;i looked for asym-
metry in upward and downward movements in the unemployment rate by examining
the matrix of Markov probabilities for transitions between these phases. 2 His use
of standard Markov chains was quite novel and was very intuitive and transparent.
With a fully specified dynamic model, Hamilton examined the same fundamen-
tal issue as Neftl;i. Hamilton demonstrated that output growth is well described by
an exogenous sequence of shifting regimes of different average growth rates and
dynamics that roughly correspond to NBER business cycle phases. Although
estimation of Hamilton's model is somewhat involved, the basic structure of the
model is very intuitive.
Beaudry and Koop tapped into the long-standing debate about persistence of
shocks to output growth. They provided evidence that negative shocks are reversed
564 MACROECONOMETIUCS
fairly quickly, while positive shocks are much more persistent. This result implies
that recessions may indeed be transitory at the same time that output is not trend
stationary. Beaudry and Koop's model for output growth is just an ARMA model
with some simple terms added that allow negative shocks-like recessions-to be
reversed fairly quickly.3
These papers succeeded because they each examined a question of importance
to economists, using straightforward techniques that avoided needless complexity.
To state this position in stronger terms, nonlinear models in and of themselves will
not capture much attention in the profession. Simply establishing that one nonlinear
model outperforms another-or outperforms a linear model-is unlikely to gen-
erate much interest, just as there is little interest in establishing whether a particu-
lar variable is best modeled as an ARMA (2, 2) or an ARMA (3, 2). In any case,
with about 200 observations in the postwar quarterly history, the true univariate
nonlinear model likely will remain elusive.
Beaudry and Koop, of course, draw out the implications of the bounceback for
the persistence of positive and negative shocks to output. 4 Sichel (1994) draws out
another important implication for macroeconomics-namely, that this bounceback
of output after recessions strongly supports the output-gap view of fluctuations of
DeLong and Summers (1988) and the plucking-model view in Friedman (1969,
1993). Both of these papers argued that output generally grows along a trend
(either stationary or nonstationary) and that fluctuations are primarily lapses be-
neath the sustainable level of production implied by the rising trend. 5 DeLong and
Summers refer to these lapses beneath trend as "output gaps," while Friedman
identifies them as "plucks." Friedman's terminology makes clear his view (which
is also implicit in DeLong and Summers) that output bounces back after it is pulled
down from its sustainable level. This pattern of contraction and bounceback is
precisely that implied by the dynamics of the floor regime in Potter's model.
Conclusion
Acknowledgments
Notes
1. There is too much research for all of it to be cited here. Potter reviews some of this literature.
Also see the references listed in Sichel (1993) and Diebold and Rudebusch (1994).
2. Neft~i provides evidence that the probability of exiting from a contraction exceeds the probabil-
ity of exiting from an expansion, implying that unemployment rises faster in a contraction than it falls
during an expansion. Sichel (1989) identified an error in Neft~i' s empirical results. However, Rothman
(1991) showed that Neft~i's original results are resurrected if a first-order Markov process is used for
unemployment rather than the second-order Markov process used in Neft~i's paper.
3. Iwata and Hess (1994) suggest that standard inference is inappropriate for the extra terms added
by Beaudry and Koop. Their results raise questions about the significance of the nonlinear terms.
4. Sichel (1994) demonstrated that the bounceback in output reflects the dynamics of inventory
investment; that is, real final sales-output less inventory investment-do not exhibit bounceback
dynamics. Furthermore, that paper shows that inventory investment and the inventory-sales ratio are
stationary variables. This fact provides further evidence that output has an important transitory com-
ponent because changes in inventory investment account for so much of the variance of output growth
at business cycle frequencies.
5. Sichel (1993) shows that output exhibits "deepness." That is, dips below trend tend to be deeper
than rises above trend. This stylized fact also supports the view implicit in Delong and Summers' and
Friedman's work.
6. As indicated earlier, Sichel (1994) showed that inventories playa key role in the nonlinear
behavior of output. French and Sichel (1993) find evidence that nonlinearities arise in durable goods
output and structures, but not in services. Rothman (1991) looking at unemployment, demonstrated
that asymmetry arises in the manufacturing sector.
7. For some examples, see Ball and Mankiw (1991), Hamilton (1988), and Cecchetti, Lam, and
Mark (1990).
8. James Cover (1992) provides evidence that output responds more to negative money shocks
than to positive money shocks.
References
Balke, N., and M. Wynne. (1992). "The Dynamics of Recoveries." Mimeo, Federal Re-
serve Bank of Dallas.
Ball, Lawrence, and Gregory Mankiw. (1991). "Asymmetric Price Adjustments and Eco-
nomic Fluctuations." Manuscript, Harvard University.
Beaudry, Paul, and Gary Koop. (1993). "Do Recessions Permanently Change Output?"
Journal of Monetary Economics, 149-163.
Cecchetti, S.G., P. Lam, and N.C. Mark. (1990). "Mean Reversion in Equilibrium Asset
Prices." American Economic Review 80, 398-418.
Cover, James P. (1992). "Asymmetric Effects of Positive and Negative Money-Supply
Shocks." Quarterly Journal of Economics November, 1261-1282.
DeLong, Bradford, and Lawrence Summers. (1988). "How Does Macroeconomic Policy
Affect Output?" Brookings Papers on Economic Activity 1988(2), 433-494.
Diebold, Francis X., and Glenn D. Rudebusch. (1994). "Measuring Business Cycles: A
Modem Perspective." Mimeo, Federal Reserve Board and University of Pennsylvania.
COMMENTARY ON CHAPTER 13 567
Friedman, Milton. (1969). "Monetary Studies of the National Bureau." In The Optimum
Quantity of Money and Other Essays (ch. 12, pp. 261-284). Chicago: Aldine.
Friedman, Milton. (1993). "The 'Plucking Model' of Business Fluctuations Revisited."
Economic Inquiry, 171-177.
French, Mark W., and Daniel E. Sichel. (1993). "Cyclical Patterns in the Variance of
Economic Activity." Journal of Business and Economic Statistics 11(1), 113-119.
Hamilton, James D. (1988). "Rational-Expectations Econometric Analysis of Changes in
Regime: An Investigation of the Tenn Structure of Interest Rates." Journal ofEconomic
Dynamics and Control 12,385-423.
Hamilton, James D. (1989). "A New Approach to the Economic Analysis of Nonstationary
Time Series and the Business Cycle." Econometrica 57, 357-384.
Iwata, Shigeru, and Gregory D. Hess. (1994). "Asymmetric persistence in GDP? A Deeper
look at Depth." Mimeo, University of Kansas.
Neft~i, Salih N. (1984). "Are Economic Time Series Asymmetric Over the Business Cycle?"
Journal of Political Economy 92(2), 307-328.
Rothman, Philip. (1991). "Further Evidence on the Asymmetric Behavior of Unemploy-
ment Rates Over the Business Cycle." Journal of Macroeconomics 13,291-298.
Sichel, Daniel E. (1989). "Are Business Cycles Asymmetric: A Correction." Journal of
Political Economy 97(5), 1255-1260.
Sichel, Daniel E. (1993). "Business Cycle Asymmetry: A Deeper Look." Economic Inquiry
31,224-236.
Sichel, Daniel E. (1994). "Inventories and the Three Phases of the Business Cycle." Journal
of Business and Economic Statistics 12(3), 269-277.
Subject Index
Aggregate demand and supply 355. 371, Bayesian econometrics 30. 35, 38-9, 64-5,
374 90
Aggregate price level 326--7 Business cycle
Aggregate productivity 367 asymmetrical business cycles 9. 533-4.
Applied macroeconomics 110. 127.457-60 547
Autoregressive conditional heteroskedasticity calibration of models 32-5. 190
(ARCH) models 9. 427-66 causes of 365-71
applications of 446-57 general equilibrium approach, history of
estimation and testing 441-6 183-91
eXpOnential GARCH (EGARCH) model endogenous delay threshold autoregression
446 534-44
fractionally integrated GARCH floor and ceiling models of U.S. output
(FIGARCH) model 457-8 535.541-4
future research on 457-60 fluctuations 191-2.356-9.517
generalized ARCH (GARCH)-in-Mean historical approach 351- 406
(GARCH-M) model 447 long-run cycles 384-9
GARCH process 433-4, 467 regime-switching model 533
integrated GARCH model (IGARCH) shocks 182. 188. 189. 193-6.365-70.
447-8 384-9.525-6.547-52,564-5
limitations of 458-60 statistical tests 199-207
multivariate GARCH models 449-50 structural approach, model estimation and
motivation for 428, 434-6 evaluation. 30-41
properties of 433-4, 436-41 system-of-equations approach 183. 184-6
Autoregressive moving average (ARMA) timing of 358-9 and VAR models 115
models 434
Asset prices 457, 460. 467, 468
Calibration of models 32-5. 115, 190 see
also Macroeconometric models
Bank failure 393-4, 403n.15 Causal structure 8, 211-27
Banking Act of 1933 397 formal characterization of. 212-18
Banking system. American, 1929-1935 application of causal orderings 215-18
390-8, 403n.12, 403n.14, 403n.16 parameter-variable distinction 219-20
569
570 SUBJECf INDEX
Shocks, macroeconomic 22, 76, 79-83, 527, Unit roots 63, 447-8, 474-80
547-52, 564-5 see also Time series Unit root vs nonunit root processes 474-80
models
Stocks 461,529-30
VAR (vector autoregression) models 6,
Stochastic volatility models 448-9, 450-1,
57-91,126
457
Bayesian VAR approach 64-5, 90
Stochastic processes and nonstationary series
and calibration approach 69-70
473-99
and causal analysis 221-3
Structural business cycle models 30-41,
estimation and specification 59-66
47-53, 59-66
evaluation of 90-1,72-3, 104
Bayesian methods 30, 35, 38-9, 64-5,
forecasting 89-90, 104
90
identifying generic model of economic
calibration approach 32-5, 69-70, 115
behavior 70-9
generalized method of moments (GMM)
interpretation of results 79-88, 104
30,32,35-7,69, 102-4, 172, 177,
and Lucas's critique 68
200,205,445
methodology 66-73, 90
maximum likelihood estimation 30-2,
parsimonious VAR 153
65
and rational expectations approach 68, 99
method of simulated moments 30, 33-4,
semistructural unrestricted VAR analyses
37-8
71-3
Structural models
shocks 72, 73, 76, 79-83, 86, 104
causal orderings 212-18
specification of VAR models 64-6
definition 66-7, 212, 223
and theoretical models 90-1, 11 5
post-Keynesian approach 17-23
Volatility 356-9
semistructural VAR analyses 57-91
and long memory 458-60
vs rational expectations models 68
and misspecification 458
Superexogeneity 266, 281-301, 307n.2, 314,
fractionally integrated GARCH
315,495
(FIGARCH) model 457-8
conditional volatility dynamics 427-66
forecasts 454-7
Taxation 396-7
multivariate models 450-4
Technology shocks 367-71
stochastic volatility models 457
Thatcher government 143
Time series models
and regime changes 161-2, 325-44, Wages, prices, and unemployment in UK
355-65 131-62
and shock estimation 39-41, 79-83, Wage rates, U.S. 352,374-7,394
366-70, 527-8 Wage rigidity 374-7
limitations of, 353, 355-89,419,421, Wald test 491
398-402 Walras law 4
linear and nonlinear, comparison 517-30, Walrasian model 4
599n.2, 563-5 Weak exogeneity 113-14, 124-6, 132-3,
nonlinear models of U.S. output 533-4, 136, 151, 152, 159, 160, 173,290,
564-5 292,474,491-3,494-5,497,501,
nonlinear stochastic models 23-30 502,506,507,509,513,514
volatility modeling 428 see also Witsenhausen's causality property 253-6,
VAR models, LSE methodology, 26On.1
Historical macroeconomics, ARCH Wold theorem 60
models Wold representation 519-23
Author Index
575
576 AUTHOR INDEX
Quah, D., 84
Wallace, N., 242, 244-6, 326, 328
Walley, 1., 188
Rabemananjara, R., 126 Weiss, A.A., 442
Richard, J.-F., 110-11, 174,298 West, K.D., 460
Robinson, P.M., 457, 46In.15-16 Wicksell, K., 193
Rodrik, D., 338 Wilcox, 1.A., 279
Rogin, L., 248-9 Williams, R., 110
Romer, C, 356, 373 Witsenhausen, H.S., 253, 254
Rowley, R., 110 Wold, H.O., 428, 519
Rudebusch, G., 346, 358 Wooldridge, J.M., 442, 449
Wymer, C., 110